Meta Muse Wonder: Benchmarks, Pricing and Access (2026)

Meta's first frontier model, tested against Claude Opus 5.5, GPT-6 Astra and the rest: what the benchmarks really show, what it costs per task, and which door gets you Muse Wonder Full

Meta priced its first frontier model to the cent at the rate of the model that beats it. On October 12, 2026, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, and the standard model went on sale through the Meta Model API at $4 input and $20 output per million tokens, with cached input at $0.20 - Meta. Those are exactly the rates Anthropic charges for Claude Opus 5.5 - Claude pricing. On Artificial Analysis's independent index, Opus 5.5 scores 58, Wonder Full 56 and the API-available Wonder 52 - Artificial Analysis. The launch post reached the top of Hacker News with more than 600 points on day one - Hacker News, and TechCrunch called the pair Meta's "first attempt at a frontier model" - TechCrunch.

The problem is that "Muse Wonder" names two different models with two different access rules, and almost every headline blurs them. The model that wins Meta's most impressive benchmarks, Wonder Full, is not on the API at all: it lives only inside the Muse Code terminal agent, on plans costing $50 or $200 a month, where Meta trains on your code by default unless you switch it off. The model you can build a product on is the standard Wonder, which ranks in a cluster with GPT-6 Astra and Gemini 4 Argon rather than at the top. And Meta's launch table, while unusually well documented, mixes independent results with benchmarks Meta built and graded itself.

This guide separates all of that. It explains what Meta actually released and how it got there in six months, walks through Meta's benchmark table row by row, sets it against independent results from Artificial Analysis and Vals AI, and calculates what Wonder costs per finished task rather than per token. It maps the four doors to the models (API, OpenRouter, Muse Code and the Muse app), covers safety and data handling, explains why the agent harness matters as much as the model, and ends with a decision guide by job. No technical background is assumed; where a section gets technical, the plain-language version comes first.

Contents

1. Muse Wonder on one page
2. What Meta shipped on October 12, and how it got here
3. Meta's benchmark table, read row by row
4. What the independent numbers say
5. How to read Meta's benchmarks without being fooled
6. Price: what Muse Wonder actually costs
7. How to get access: the four doors
8. Muse Wonder Full and the Muse Code plans
9. Safety, privacy and keeping an agent in bounds
10. The harness matters as much as the model
11. The decision guide: when to use Muse Wonder
12. Where Muse Wonder falls short
13. The outlook
14. The bottom line

Muse Wonder against the frontier: the scored comparison

Before the detail, here is where both Wonder models land among the options a buyer realistically weighs this month. The table scores nine models on four criteria using independent data wherever it exists, so that Meta's models are judged by the same evidence as everyone else's. Capability, agent work and cost come from Artificial Analysis (Intelligence Index v4.3, Terminal-Bench 4.0, AutomationBench-AA and cost per index task, retrieved October 13, 2026), and access is scored from each vendor's own documentation - Artificial Analysis. Each cell shows the score and the measurement behind it.

How the criteria work. Capability (30%) maps the Artificial Analysis Intelligence Index linearly from 40 (a score of 0) to 58 (a score of 10), because the index is the broadest independent measure and small gaps compound in long agent runs. Agent work (20%) averages two Artificial Analysis-run agent tests, Terminal-Bench 4.0 (terminal coding) and AutomationBench-AA strict completion (business workflows, a Zapier-built benchmark run by Artificial Analysis), and maps the average linearly from 30 (0) to 56 (10) - Artificial Analysis. Cost per task (25%) uses Artificial Analysis's weighted cost per index task on a log scale from $0.25 (10) to $10 (0), because buyers think in multiples. Access (25%) scores how easily a business can use the model today: self-serve API, third-party routing, subscription coverage, data modes and availability. Muse Wonder Full has no per-task cost, so, following our standard method, that criterion is excluded from its average rather than guessed. Final scores are weighted averages rounded half up, and ties are ordered alphabetically.

Read the table as a buyer's default, not a verdict on capability alone. It says three things about Meta's launch. First, Wonder Full is genuinely frontier-class on capability and agent work, scoring 8.8 and 8.2, but its access score pulls it to fifth, because a model you cannot call from your own software is worth less to a business than one you can. Second, the standard Muse Wonder is a mid-pack choice on value: same token price as Opus 5.5, lower capability, and a per-task cost that beats Anthropic's top models but loses to OpenAI's. Third, the top of the table belongs to models that are not the most capable but deliver the most per dollar, which is why the rest of this guide spends as much time on cost and access as on benchmarks. If your job happens to be one where Wonder's specific strengths matter (long-horizon coding, low hallucination, legal and workflow deliverables), section 11 shows where it moves up this ranking.

1. Muse Wonder on one page

Muse Wonder is a pair of large reasoning models from Meta Superintelligence Labs, released on Monday, October 12, 2026: Muse Wonder and the higher-capability Muse Wonder Full. Meta calls them "our most capable models" and frames them as "the next step on our scaling ladder and a major leap toward personal superintelligence" - Meta. TechCrunch described the pair as "the company's first attempt at a frontier model" - TechCrunch, which is the clearest one-line summary of why the launch matters: for the first time since Meta rebuilt its AI effort, it has a model that sits near the very top of the independent rankings rather than in the pack behind them.

Both models take text, images, video, audio and PDF documents as input and return text. Both have a context window of 1,048,576 tokens and support six reasoning effort settings from "minimal" to "max" - Meta Model API docs. They were trained at Meta's gigawatt-scale Prometheus data center, and Meta says they build on "larger, stronger base models" than the Muse Spark family that preceded them. Neither model's weights are downloadable.

The spec sheet below collects the facts a buyer needs in one place, each drawn from Meta's own documentation or model card.

Source: Meta Model API docs, with prices, plans and benchmark scores from the Meta pages cited in sections 3, 6 and 8.

Three facts from that table shape everything that follows. First, the two variants are distinct models, not two settings of one. Meta's docs state that setting reasoning effort to "max" on Muse Wonder "doesn't select" Wonder Full; they are distinct variants with different scores. Second, the stronger model is harder to get. Wonder Full is available only to Muse Code subscribers on the two upper plans, so a developer building a product on Meta's API is building on the standard Wonder, which scores four points lower on the independent index. Third, the price is a deliberate echo: Muse Wonder's API rates match Anthropic's Claude Opus 5.5 to the cent, which section 6 unpacks.

1.1 What Meta says the models are for

Meta pitches the Wonder series around two named capabilities. Active Memory is "carrying important context about you across conversations and actively retrieving the details relevant to the task at hand"; Active Focus is how the model "continues working toward a challenging goal: trying alternative approaches, working through blockers, and adapting to new information", while asking for clarification and checking in before important actions - Meta. On its developer site, Meta translates those into four practical claims: frontier coding in large repositories, long-running agents that "run for hours, uninterrupted" while delegating to sub-agents, advanced computer use across browsers, desktop and mobile apps, and memory that persists through conversations - Meta.

Meta's official launch video is the quickest way to hear the pitch from the researchers who built the models. It runs about two and a half minutes and walks through the Active Memory and Active Focus ideas with demos of personal and professional tasks.

The memory theme is the one the researchers lean on hardest. "We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months, and beyond," research scientist Kavya Kopparapu says in the video - TechCrunch. Section 5 explains why the benchmark behind that claim deserves caution.

That positioning is narrower and more agent-focused than a typical flagship launch. Meta is not claiming the best chatbot or the best general reasoner. It is claiming the best model for agents that act on a person's behalf over long stretches of time, at work and at home, and it chose its benchmarks accordingly. That choice is the key to reading the rest of this guide: Wonder looks strongest exactly where Meta aimed it, and ordinary where it did not.

Why this matters: "Muse Wonder" in a headline can mean either of two models with different scores and different access rules, and most coverage blurs them. How to apply this: whenever you see a Muse Wonder benchmark, check which variant it refers to and whether you can actually use that variant through the door you plan to use.

2. What Meta shipped on October 12, and how it got here

Muse Wonder is the end point of a fast and expensive rebuild. Meta's launch post opens its history in one line: "Fourteen months ago, we began a ground-up overhaul of our artificial intelligence efforts. In that time, we've gone from rebuilding our stack from first principles to pushing the frontier" - Meta. The overhaul began when Mark Zuckerberg restructured Meta's AI organization in August 2025 and brought in Scale AI founder Alexandr Wang to lead what became Meta Superintelligence Labs, after a disappointing developer reception for Llama 4 - CNBC. Zuckerberg marked the anniversary in his own launch post: "It's been just over a year since we started Meta Superintelligence Labs" - Mark Zuckerberg on Threads.

Understanding that path explains why this launch is framed the way it is. Meta did not arrive at Muse Wonder by scaling up Llama. It built a new model family, closed rather than open, and released it in a deliberate sequence of smaller steps before attempting the frontier. Wonder is the first step in that sequence that Meta itself, and independent evaluators, describe as frontier-class.

2.1 Six months, six models

The Muse family began on April 8, 2026 with Muse Spark, which Meta introduced as "the first in the Muse family of models developed by Meta Superintelligence Labs" and released through a private API preview - Meta. It was Meta's first major proprietary model, a break with the downloadable Llama line, and our guide to Muse Spark covers that strategic shift in detail. Spark was followed by point releases on July 9, August 5 and September 2, the last being Muse Spark 1.3 - Meta.

Around those point releases came the Muse Code terminal agent in preview on August 5, the open-weight Muse Glimmer on August 10 (a 30-billion-parameter model distilled from Spark and released under Apache 2.0) - Meta, and the consumer Muse agent app on September 8.

Meta's own launch chart, built on Artificial Analysis data, shows what that cadence bought. On the current version of the index, the original Spark scores an estimated 31, Spark 1.1 scores 34, Spark 1.2 40, Spark 1.3 48, Muse Wonder 52 and Wonder Full 56 - Meta.

A 25-point climb in six months is the strongest evidence for Meta's "predictable, rigorous scaling" claim. It also has a mundane component worth stating: the starting point was low. The original Spark sat about 27 points below today's leader on the same scale, and much of the climb was closing a gap to scores the leading labs had already reached. The last two steps, from 48 to 56, are the ones that put Meta near the top, and they came from a new and larger base model rather than from more tuning of the old one.

2.2 What is new in Wonder

Meta says little about Wonder's architecture. The model card discloses no parameter count, no architecture details and no training-data cutoff - Meta model card. What it does say is that the Wonder models "build on larger, stronger base models" than Muse Spark and combine large-scale pretraining with reinforcement learning, and Zuckerberg described the series as "a much larger and more advanced class of models", splitting it into "Muse Wonder for fast frontier performance and Muse Wonder Full for the full version of the model" - Mark Zuckerberg on Threads. Whether Wonder Full is a bigger network or the same network given more compute at inference time, Meta has not said. Internally, CNBC reports, the project was codenamed Watermelon - CNBC.

Both models were trained at Prometheus, Meta's gigawatt-scale data center in New Albany, Ohio, which TechCrunch notes came online earlier this year - TechCrunch.

Prometheus is the first of Meta's multi-gigawatt AI campuses; a second, Hyperion in Louisiana, is planned to reach 2 gigawatts by 2030 and scale toward 5 - TechCrunch.

The money behind them is on the public record: in July, Meta narrowed its 2026 capital expenditure guidance to $130 billion to $145 billion - Meta investor relations. Our Nebius guide covers how Meta has also been renting outside capacity to keep up.

Alexandr Wang's message on launch day was unambiguous: "These are true frontier models. We really think they are competitive with the best models out there" - Quartz. The independent numbers in section 4 support the first sentence for Wonder Full and qualify the second: competitive, yes; best overall, not on any independent composite yet.

2.3 Why Meta needed a frontier model at all

The first-principles question is why a company that earns its money from advertising and, increasingly, commerce would spend this much to sit at the top of a model leaderboard. The answer lies in what Meta is trying to sell next. Its stated goal is "personal superintelligence": agents that act on behalf of billions of people across messages, calendars, shopping and work. That product depends on the model in a way a chatbot does not. An agent that books, buys and negotiates for you is only as trustworthy as the reasoning underneath it, and a company that rents that reasoning from a rival hands the rival control over its most important product's quality, cost and roadmap.

Seen this way, Muse Wonder is less a product than an input. Meta says so directly: Wonder "will also power artifact generation and processes that require more intelligence", and "in the future, we'll be using Muse Wonder to develop specialized models that we can bring to everyone using Muse" - Meta. That second sentence describes distillation: using the large model to teach smaller, cheaper ones that can run at Meta's scale. A frontier model is the teacher every cheaper model needs, which is why owning one matters even if most users never touch it directly.

Why this matters: Wonder is the product of a deliberate ladder, and Meta has signaled more rungs above it, so judging it as a one-off release misses the trajectory. How to apply this: if you build on Meta's models, plan for frequent version changes (six models in six months so far), pin model IDs in production, and re-evaluate at each release rather than assuming today's ranking will hold.

3. Meta's benchmark table, read row by row

Meta published one master table with fourteen rows (thirteen benchmarks plus the Artificial Analysis index) and five models: Muse Wonder Full, Muse Wonder, Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5, all at maximum reasoning effort - Meta. It is the most-shared image of the launch and the source of almost every headline. It is also easy to misread, because the rows mix internal evaluations, third-party results and provider-reported numbers, and because "leads" on one row can mean a margin of 0.1 points or of 11.

The full table is reproduced below exactly as Meta published it, with the best score in each row in bold. The right way to read it is by row type: what is being measured, who ran it, and how large the gap is. A model that wins by a large margin on a third-party benchmark is telling you something; a model that wins by a fraction of a point on its own internal benchmark is telling you much less.

Source: Meta Muse Wonder model page, October 12, 2026. Wonder Full's SWE-Marathon score was run at xhigh effort; OSWorld binary (strict) scores are 39.4, 34.3, 43.1 and 34.3, with no binary figure for Opus 5.5.

Count the bold cells and a clearer picture emerges than the launch coverage suggested. Meta's two models take seven of fourteen rows (six for Wonder Full, one for Wonder on MemoryBench), Claude Opus 5.5 takes six, GPT-6 Astra takes one, and Claude Fable 5.1 takes none. Opus 5.5 wins the composite index, terminal coding, computer use, health, from-scratch programming and (by a tenth of a point) JobBench. That is not the profile of a model that has been dethroned. It is the profile of two strong models with different specialties.

3.1 Where Wonder Full leads, and by how much

The margins matter more than the count. The chart below sets Wonder Full against the two strongest rivals on the four rows where its lead is clearest and the three where it trails most, using the same published numbers.

The biggest gap is SWE-Marathon, where Wonder Full scores 57.5 against Opus 5.5's 46.3, an 11.2-point lead on a benchmark of "substantial projects graded by executing project tests" - Meta model card. That is the single strongest claim in the launch, and it lines up with what Meta says it trained for: "diverse, long-running tasks". Two caveats from the model card itself belong next to it. Wonder Full's score was run at xhigh rather than max effort, and the rival scores come from the leaderboard run by Abundant AI, where GPT-6 Astra ran inside Codex and Fable 5.1 inside Claude Code, "so this row is not a controlled common-harness comparison". In other words, part of the gap may be the harness, not the model.

DeepSWE 1.1 is the more controlled coding win: 79.0 for Wonder Full against 74.2 for Opus 5.5, with Datacurve as the third-party source for Meta's models and Opus 5.5, and OpenAI's own figure for Astra. A five-point lead on a 113-task, 91-repository benchmark graded by handwritten tests is meaningful, with one asterisk: Google reported 77.9% on DeepSWE for its Gemini 4 Argon two weeks earlier, calling it a new state of the art - Google, so Wonder Full's lead is over the models in Meta's table rather than over the whole field. AutomationBench-AA, Artificial Analysis's simulated business-workflow test, gives Wonder Full a four-point lead under a strict all-assertions-pass rule, and because Artificial Analysis ran it, it is among the most trustworthy rows in the table. The lead is over the five models in Meta's table, though: on Artificial Analysis's full leaderboard, Wonder Full ranks third on this benchmark, behind models Meta did not include - Artificial Analysis.

3.2 Harvey LAB-AA: a big lead in a race nobody has finished

Harvey LAB-AA is the row Meta leads its domain-expertise section with, and the one where its lead looks largest in relative terms: 13.1 for Wonder Full against 4.2 for Opus 5.5, roughly three times higher. It is Artificial Analysis's version of Harvey's Legal Agent Benchmark, in which agents produce legal work products from matter documents, and the score is the share of tasks whose deliverable "meets every criterion and passes the hallucination checks" - Meta model card.

Read the absolute number before the ratio. A score of 13.1 means that on roughly 87% of tasks, even the best model produced a deliverable that failed at least one criterion or one hallucination check. All-pass scoring is deliberately brutal (a brief that is excellent except for one invented citation scores zero), which is why everyone's numbers are low. So the honest reading is that Wonder Full fails less often than its rivals at producing a fully clean legal work product, not that it does legal work well unsupervised. Artificial Analysis's own leaderboard confirms the ranking (Wonder Full first, Wonder second, Grok 4.7 third at 9.4%, Opus 5.5 ninth) and adds a useful detail: Wonder Full still averaged 0.62 material hallucinations per task - Artificial Analysis. Section 4 shows that a different implementation of the same Harvey benchmark ranks the models quite differently. Our guide to why AI coding benchmarks mislead explains the same trap in software: a relative lead on a benchmark with a low ceiling says little about readiness for real work.

3.3 Where Wonder Full trails

The rows Meta does not lead are just as informative, because they are the ones a buyer is most likely to hit in everyday use. On Terminal-Bench 4.0, the Artificial Analysis-run test of hard tasks in isolated terminal environments, Wonder Full scores 57.1 against 59.6 for Opus 5.5 and 59.1 for GPT-6 Astra. Terminal work is the bread and butter of coding agents, so a 2.5-point deficit there tempers the long-horizon wins. On ProgramBench, which asks an agent to reconstruct a complete program in a clean-room environment with every model using the same mini-swe-agent harness, Opus 5.5 scores 70.5 against Wonder Full's 59.0, an 11.5-point gap that is the mirror image of SWE-Marathon.

That pairing is worth dwelling on. On the one coding row where Meta controlled the harness for every model (ProgramBench), Opus 5.5 wins by about as much as Wonder Full wins on the row where the harness was not controlled (SWE-Marathon). Neither row settles the question, but together they argue against reading the table as a clean coding victory for Meta. On HealthBench Pro, both Opus 5.5 (65.6) and Astra (65.3) lead Wonder Full (62.3), and on OSWorld V2.1, the computer-use test, Opus 5.5's partial score of 81.8 and Fable 5.1's 80.7 sit above Wonder Full's 78.4 (Fable also leads the strict binary score, 43.1 to 39.4).

Why this matters: the table supports a narrower claim than "Meta now leads the frontier". It supports this: Wonder Full is the strongest of these five on long, multi-step agent and coding work and on some professional deliverables, while Opus 5.5 remains stronger on terminal tasks, computer use, from-scratch programming and the broadest composite measure. How to apply this: match the row to your workload. If your agents run for hours on one large project, the SWE-Marathon and DeepSWE rows are your evidence. If they run many short terminal tasks, Terminal-Bench is, and it points elsewhere.

4. What the independent numbers say

Meta's table is one view of Muse Wonder. The more useful view comes from evaluators who choose their own tests, run every model the same way and publish results for models Meta did not pick. Within a day of launch, two of them had published results from three angles: Artificial Analysis, which runs the most-cited composite index and a separate index that scores a model inside the coding agent built for it, and Vals AI, which runs domain benchmarks in finance, law and other professions. Together they confirm the outline of Meta's story and correct its emphasis.

The headline confirmation is that Meta did not inflate the independent numbers it reused. Its scores for the Artificial Analysis index, Terminal-Bench 4.0 and Harvey LAB-AA match Artificial Analysis's own figures exactly. The correction is that Meta's comparisons left out several models that change the ranking, and that on some benchmarks Meta reports as wins, the full independent leaderboard looks different.

4.1 The Artificial Analysis Intelligence Index

The Artificial Analysis Intelligence Index v4.3 combines ten evaluations (AA-Briefcase, GDPval-AA, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR), weighted 30% agents, 20% coding, 20% scientific reasoning and 30% general knowledge - Artificial Analysis methodology.

On that index, Artificial Analysis scored Muse Wonder Full at 56 and Muse Wonder at 52, both at max reasoning effort, "each a material gain over Muse Spark 1.3 (max) at 48" - Artificial Analysis.

The chart shows how compressed the top of the market has become. Eight models from four labs sit within six points of each other, and the order changes depending on decimal places: on Artificial Analysis's unrounded figures, Claude Sonnet 5.5 (56.00) edges Muse Wonder Full (55.80), which is why Meta's launch chart, which omits Sonnet 5.5, shows Wonder Full in second place while the full leaderboard shows it third - Artificial Analysis. The standard Muse Wonder, the API model, sits in a cluster with GPT-6 Astra, Gemini 4 Argon and GPT-6.1 Sol at around 52 to 53.

Underneath the composite, Artificial Analysis reports where Wonder Full ranks on individual evaluations: first on CritPt (33%) and GDP.pdf (33%), second on GDPval-AA (an Elo of 1857), and third on AA-Briefcase, AutomationBench-AA (71%) and SciCode (63%) - Artificial Analysis. Those are strong placements on exactly the professional-work evaluations Meta emphasized.

Its weakest areas are factual recall and terminal coding: on AA-Omniscience, which rewards correct answers and penalizes confident wrong ones, Wonder Full scores 30 against 46 for Opus 5.5 - Artificial Analysis. The flip side is one of the most underreported results of the launch. Wonder Full's hallucination rate (how often it answers wrongly instead of declining when it does not know) fell from Muse Spark's 33% to 23%, the second-lowest of any model scoring 45 or more on the index, behind only Gemini 4 Argon at about 15%.

Artificial Analysis's model data puts Claude Opus 5.5 at roughly 59% and GPT-6 Astra at roughly 51% on the same measure - Artificial Analysis. In plain terms, Wonder knows fewer facts than Opus but is much more willing to say when it does not, which is a valuable trait for agents that act on what they believe.

4.2 Speed and token use

Artificial Analysis also measures how fast and how wordy each model is, which matters for agents that run thousands of steps. Muse Wonder on Meta's own API produced a median of about 101 output tokens per second, with roughly 20 seconds to the first token and about 39 seconds to the first answer token at max effort - Artificial Analysis. That is in line with Claude Opus 5.5 (about 95 tokens per second) and faster than GPT-6 Astra (about 63). Wonder Full has no speed data, because it has no per-token API to measure.

Token use is where Wonder's cost profile comes from. Both Wonder models use around 54,000 output tokens per Intelligence Index task, but the bigger number is on the input side: agentic evaluations reread large contexts many times, and input accounts for $3.43 (76%) of Muse Wonder's $4.52 cost per task - Artificial Analysis. The practical lesson is that Wonder's bill on agent work is driven more by how much context it rereads than by how much it writes, which makes the $0.20 cached-input rate the most important number on its price card.

4.3 Coding agents: the model inside its harness

Artificial Analysis's Coding Agents Index scores a model inside the agent product built for it, averaging DeepSWE, Terminal-Bench 4.0 and SWE-Atlas-QnA. Here the result is genuinely close at the top: Claude Code with Sonnet 5.5 scores 68.4, Muse Code with Wonder Full 68.0, Claude Code with Opus 5.5 66.0, Google's Antigravity CLI with Gemini 4 Argon 63.8, and Muse Code with the standard Wonder 63.5 - Artificial Analysis.

This is the most decision-relevant independent result for developers, because it measures what you would actually run. It says Muse Code with Wonder Full is a top-tier coding agent, level with Anthropic's best, with a clear edge on long-horizon DeepSWE tasks (77.0) and a clear deficit on Terminal-Bench (59.1 against 66.2 for Claude Code with Sonnet 5.5). It also says the standard Wonder, the one available on the API, is a tier below. Section 10 returns to why the harness matters as much as the model.

4.4 Vals AI and the professional benchmarks

Vals AI ranks Muse Wonder Full fifth of 47 models on its composite Vals Index (64.57%) and the standard Wonder tenth (61.06%), with Gemini 4 Argon leading at 68.90%; Vals also ranks Muse Wonder first of 47 on its CyberBench (80.36%) - Vals AI.

More pointedly, two of the professional benchmarks Meta presents as wins look different on Vals' own leaderboards. On Vals Finance Agent V2, where Meta's table shows Wonder Full leading at 59.1, the full Vals leaderboard ranks it fifth (59.06%), behind Gemini 4 Argon (65.40%), Gemini 3.8 Flash and two of Meta's own Muse Spark models - Vals AI.

On Vals' separate implementation of Harvey's Legal Agent Benchmark, Wonder Full ranks fourth at 21.67%, behind Muse Spark 1.2 (25.42%) and both Muse Spark 1.3 variants - Vals AI. That is a striking result: on one legal benchmark Meta's newest model leads everyone (Artificial Analysis's version), and on another implementation of the same underlying benchmark Meta's older, cheaper models beat it. It is a reminder that "legal agent performance" is not one number, and that implementation details (task sets, judges, hallucination rules) can reorder a leaderboard.

4.5 What has not been measured yet

Several independent sources had not yet covered Muse Wonder at the time of writing. The LMArena text leaderboard, which ranks models by human preference votes, did not list it (its leader was Gemini 4 Argon) - LMArena, and Epoch AI's capabilities index did not list it either, with Claude Opus 5.5 at the top - Epoch AI. The official Terminal-Bench leaderboard, led by Opus 5.5 inside Claude Code at 64.8%, had no Muse Wonder entry - Terminal-Bench. These gaps will close over the coming weeks, and they are worth rechecking before any large commitment.

One more caution applies to every number in this section: index versions are not comparable across time. Muse Spark 1.3 scored 62 on the Artificial Analysis index at its September launch and scores 48 on today's v4.3, because the index itself was rebuilt with harder evaluations - Artificial Analysis. Any comparison that mixes an old score with a new one is meaningless.

4.6 Early hands-on reactions

Benchmarks are one kind of evidence; developers trying the model on their own work are another, noisier kind. The Hacker News launch thread, which passed 600 points and 300 comments, is a reasonable first sample, with the obvious caveat that a day of anecdotes is not a measurement. The most common reaction was respect for the scores combined with doubt about everyday quality. One developer who ran an image-to-HTML test reported that Wonder was "much worse than Opus 5.5", closer to Sonnet or Haiku on that task - Hacker News, which is consistent with the visual-reasoning limitation Meta's own model card acknowledges.

Another, after a broad suite of real development work, judged it "still a step below Astra/Opus5.5/Fable", adding that "intelligence is there I think, but 'taste' and 'judgment' is not there yet" - Hacker News. That matches the shape of the independent data: strong on long structured tasks, less convincing on the open-ended judgment calls that benchmarks rarely capture.

Usage limits drew comment too. A Muse Code subscriber on the $15 High Usage plan reported that two small prompts to Wonder used 4% of a weekly quota, prompting a switch back to the cheaper Spark model - Hacker News. Treat all of these as signals of where to look in your own testing, not as verdicts: a day of anecdotes from enthusiasts says more about first impressions than about sustained performance.

Why this matters: independent evaluators confirm that Muse Wonder Full is a genuine frontier model, within a whisker of the best on the broad index and level with Anthropic's best inside a coding agent, while the API-available Wonder is a strong second-tier frontier model. How to apply this: use the Artificial Analysis index and the Coding Agents Index as your neutral baseline, treat Meta's domain wins as hypotheses to test, and remember that the model you can call through the API is the 52, not the 56.

5. How to read Meta's benchmarks without being fooled

Every lab's launch table is a piece of persuasion as well as a piece of measurement, and Meta's is better documented than most. Its model card describes, benchmark by benchmark, who ran each test, which harness each model used and where Meta departed from the official setup. That transparency is genuinely useful, because it lets a careful reader separate the claims that are hard to argue with from the ones that rest on Meta's own choices. This section does that separation, so you can decide how much weight each headline deserves.

The starting point is a structural fact about launch tables: the company choosing the benchmarks also chooses which of its strengths to show. That is not dishonest, and Meta is not unusual in doing it. But it means the table answers "where is this model strong?" far better than "is this model better than the others?". The second question needs benchmarks the lab did not pick, run by people the lab does not employ.

5.1 Four kinds of number in one table

The thirteen benchmarks in Meta's table come from four different kinds of source, and they deserve different levels of trust. The model card spells out the source for each one - Meta model card.

The distinction matters because each kind of source fails in a different way. An independent evaluator can still pick a narrow test, but it cannot quietly tune the setup for one vendor. A lab running a public benchmark in its own framework keeps the official tasks and graders but controls the prompts, tools and retries. A lab-built benchmark measures what the lab chose to value, graded the way the lab chose to grade it. And a provider-reported rival score was produced under conditions nobody else can see. Sorting the rows this way takes a few minutes and changes how the table reads:

• Independent, third-party run: the Artificial Analysis Intelligence Index, Terminal-Bench 4.0-AA, AutomationBench-AA and Harvey LAB-AA (all run by Artificial Analysis), Vals Finance Agent V2 (Vals AI), GDP.pdf (Surge), DeepSWE (Datacurve, for most models) and SWE-Marathon (Abundant AI).
• Meta-run on public benchmarks: HealthBench Pro, OSWorld V2.1 and ProgramBench, run "in a common internal framework" with official tasks and graders.
• Meta-built and Meta-run: MemoryBench, FocusBench, the Verifiable Browsing Task and the Personal Superintelligence Index composite.
• Provider-reported for rivals: several competitor figures come from Anthropic's and OpenAI's own system cards or reports rather than from a common run.

Sorted this way, the independent rows tell a more balanced story than the headline. Of the eight independently run rows, Wonder Full leads five (Harvey LAB-AA, Vals Finance, AutomationBench, DeepSWE, SWE-Marathon) and trails on three (the AA index, Terminal-Bench and GDP.pdf), and one of the five leads (Vals Finance, 59.1 against Fable 5.1's 58.9) is within any reasonable margin of error. On Vals' full leaderboard for that benchmark, Wonder Full ranks fifth, behind Gemini 4 Argon, Gemini 3.8 Flash and two of Meta's own Spark models (section 4.4). In fairness to Meta, the selection also cuts the other way once: on GDP.pdf, Meta reported Surge's figures, where GPT-6 Astra leads, even though Artificial Analysis's run of the same benchmark ranks Wonder Full first - Artificial Analysis. The rows built and run entirely by Meta are where the most marketable claims live, including the Personal Superintelligence Index, which Meta defines as "an internal composite of seven personal and knowledge-work evaluations", two of which (FocusBench and VBT) are Meta's own internal tests. A model winning its maker's composite index is the least surprising result in AI.

5.2 The MemoryBench naming problem

MemoryBench shows why sources matter. Meta's model card describes it as "an internal benchmark" that tests recall across multi-turn conversations of 64K to 100K tokens, with distractor conversations and injected turns that update earlier facts, scored by an LLM judge against weighted rubrics - Meta model card. TechCrunch's launch story, however, linked the name to a public academic benchmark called MemoryBench, a paper about continual learning and memory construction frameworks - TechCrunch. Those are not the same test.

The confusion matters because "Muse Wonder leads MemoryBench" sounds like an externally verified result, and it is not one. It is a result on a benchmark Meta designed, ran and graded with a model judge. There is also a detail in Meta's own numbers that the coverage skipped: the standard Muse Wonder (85.6) scores slightly higher than Wonder Full (85.5) on it. That flat result between a smaller and a larger model suggests either that memory at this context length is saturating, or that the benchmark is not very sensitive to model scale. Either way, Active Memory is a promising design goal, not yet a proven differentiator. Our guide to AI agent memory architectures explains why remembering facts across sessions is as much a system design problem as a model property.

5.3 Score selection and the harness question

Two sentences in the methodology section deserve to be read slowly. The first: where no comparable third-party result exists, Meta uses "the highest of our internally validated run, self-reported score, or official leaderboard score". The second: "Internally generated competitor runs use common settings where practical but remain best-effort: prompts, tools, and runtime may not be optimally tuned for proprietary models" - Meta model card. Meta is being candid here, and the candor cuts both ways. Taking the highest available score for each model is fair to rivals in principle, but best-effort runs of rival models inside Meta's framework are likely to understate those rivals, because every model performs best in the harness it was tuned for.

That last point is not abstract: Meta's own card lists the agent harness among the factors that change Muse Wonder's performance, and section 10 shows a documented twenty-point swing for one model across harnesses, larger than most gaps in this table.

5.4 What Meta compared against, and what it left out

A launch table is also defined by its columns. Meta compared against Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra, three of the strongest models available, which is a credible choice. But two omissions are worth noting. Meta's own Artificial Analysis chart includes Google's Gemini 4 Argon (53) and OpenAI's GPT-6.1 Sol (52) but not Claude Sonnet 5.5, which ties Wonder Full at 56 on the index (and edges it on unrounded figures, as section 4 shows) at half of Wonder's per-token price - VentureBeat. And the safety charts in Meta's launch post compare the Wonder models against Fable 5.1 and GPT-6 Astra, but not against Opus 5.5, the model that leads the capability table.

Meta's own chart, reproduced below, shows exactly which models it chose to compare on the index. Every bar is an accurate Artificial Analysis score; the question is which bars are not there.

Neither omission is proof of anything, and there are innocent explanations (Sonnet 5.5 is a mid-tier model, and safety evaluations are expensive to run). But a buyer should notice that the two comparisons most likely to complicate Meta's story are the two that are absent: a cheaper Anthropic model with the same index score, and the strongest Anthropic model on the safety measures.

5.5 Meta's own argument about "benchmaxxing"

To its credit, Meta's launch post addresses the obvious objection head-on. It acknowledges that complaints of "benchmaxxing" and praise for "big model smell" both point at the same thing, "a breadth of capability that popular benchmarks in a few domains can't fully capture", and then offers evidence that Wonder Full generalizes better than its predecessor - Meta.

The evidence is the chart below. Each point is one reasoning effort; the horizontal axis is the score on tasks Meta targeted during post-training ("in-distribution"), and the vertical axis is the average score on tasks it did not target ("out-of-distribution").

At any given in-distribution score, Wonder Full's points sit roughly ten percentage points higher on unfamiliar tasks than Spark 1.3's trend line predicts, and Meta attributes the shift to "larger, stronger base models". This is a thoughtful piece of evidence, and it is the right kind of argument to make. Its limits are that both axes are composites of Meta-chosen tasks, the comparison is only against Meta's previous model rather than against rivals, and the out-of-distribution ceiling is still under 50%. It shows that Wonder Full generalizes better than Muse Spark, which is what a bigger base model should do. It does not show that Wonder Full generalizes better than Opus 5.5.

Why this matters: most of Meta's strongest claims hold up on independent rows, but the most quotable ones (the composite index, memory, focus) rest on Meta's own tests. How to apply this: give independent rows full weight, give Meta-run public benchmarks partial weight, treat Meta-built benchmarks as statements of intent, and run your own evaluation before committing a production workload. Our guide to agent evals and benchmarks covers how to build a small internal test set that answers your question instead of the vendor's.

6. Price: what Muse Wonder actually costs

Muse Wonder's list price is easy to state and easy to misjudge. On the Meta Model API it costs $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, with the same rate across its entire 1,048,576-token context window, plus $2.50 per 1,000 web searches when you use the built-in search tool - Meta pricing docs. "No long-context premium", in Meta's words, is a real advantage for agents that carry large contexts, since some providers charge more once a prompt passes a size threshold.

What makes this price interesting is not its size but its exact match. Anthropic charges $4 input, $0.20 cache read and $20 output for Claude Opus 5.5 - Claude pricing. Meta priced its first frontier model to the cent at the level of the model that beats it on the broadest independent index. That is a positioning statement. Pricing below Opus would have signaled a second-tier product; pricing above it would have been hard to defend with a lower index score. Pricing at par says "judge us as equals", and it means the comparison buyers make will come down to capability per task and cost per task rather than to the rate card.

6.1 The frontier price sheet

The table below sets Muse Wonder against the models it is most often compared with, using official or primary list prices per million tokens. It shows Meta pricing at the top of the "standard frontier" band and well below the premium flagships.

Sources: Meta's and Anthropic's price pages cited above; OpenAI model prices as listed in OpenRouter's model catalog, October 13, 2026.

The chart makes the market structure visible. There is a $50 premium tier (Astra, Fable 5.1), a $20 flagship tier where Wonder and Opus 5.5 sit together, a $10 workhorse tier (Sonnet 5.5, GPT-6.1 Sol) and a sub-$5 efficiency tier where Meta's own Muse Spark lives. Wonder's position means it is two and a half times cheaper than Astra on paper and twice as expensive as Sonnet 5.5, which matters because Sonnet 5.5 scores the same as Wonder Full on Artificial Analysis's index (section 4). If OpenAI's $10 tier is on your shortlist, our guide to choosing between GPT-6 Sol, Astra and Luna breaks down that family in the same way.

6.2 Why price per token misleads

Price per token is what you pay for each unit of text; what you actually care about is what you pay per finished task. The two diverge because models differ in how many tokens they spend, especially reasoning tokens, which Meta bills at the full output rate. A model with a lower token price can still cost more per task if it thinks longer, and a model with a higher token price can cost less if it gets to the answer in fewer steps.

Artificial Analysis measures this directly by recording what it costs to run each model through its full evaluation suite. Its estimate for the average Intelligence Index task puts Muse Wonder at $4.52, below Claude Fable 5.1 ($7.63), Claude Opus 5.5 ($5.98) and Claude Sonnet 5.5 ($5.46), and above GPT-6 Astra ($3.26), Gemini 4 Argon ($1.99) and GPT-6.1 Sol ($0.72) - Artificial Analysis. Two things follow. Wonder is about 24% cheaper per task than Opus 5.5 despite identical token prices, which means it uses fewer tokens to finish the same work; and GPT-6 Astra, with token prices two and a half times higher, is cheaper per task than either, which means it gets through Artificial Analysis's tasks with far fewer tokens. On a per-task basis, the rate card ranking is almost upside down.

The chart also explains where Wonder's money goes. Artificial Analysis found that input accounts for 76% of Wonder's cost per task, because agentic evaluations reread large contexts many times, while output stays around 54,000 tokens per task. The same source notes that Wonder costs 2.8 times as much per task as Meta's own Muse Spark 1.3, which scores four points lower on the index; for many workloads, that older and cheaper model is the better buy.

That is why every serious cost comparison in this guide uses cost per task. Our GPT-6 Astra pricing analysis shows the same inversion for OpenAI's flagship, and our guide to prompt caching shows how the cached-input rate (here $0.20, a 95% discount on fresh input) can matter more than either headline price for agents that resend the same long context on every turn.

6.3 A worked example

Consider a coding agent that runs 1,000 tasks a month, each reading 150,000 tokens of repository context (80% of it cached after the first turn) and writing 20,000 tokens of reasoning and output. On Muse Wonder that is 30,000 fresh input tokens at $4 per million ($0.12), 120,000 cached tokens at $0.20 per million ($0.024) and 20,000 output tokens at $20 per million ($0.40), or roughly $0.54 per task and $544 a month. The same token profile costs exactly the same on Claude Opus 5.5, because the rates are identical, and about $1.42 per task on GPT-6 Astra at its list prices.

The example is deliberately simple, and it hides the variable that matters most: different models will not produce the same 20,000 output tokens for the same task. If Wonder needs 15,000 tokens where Opus needs 20,000, as Artificial Analysis's per-task figures suggest it often does, Wonder's monthly bill falls to about $444 while Opus stays at $544. Astra is the instructive case: even if it needed only 8,000 output tokens, its bill would still be about $820, because its input and cached-input rates ($10 and $1 per million) dominate on a context-heavy job like this one. Astra's per-task advantage on Artificial Analysis's suite comes from consuming fewer tokens of every kind; on a job where the amount of context is fixed by the task (the agent must reread the same codebase whichever model runs it), that advantage shrinks or reverses. The input-to-output mix of your workload decides which model is cheapest, so the only way to know is to run your tasks and read the usage object, which on Meta's API reports reasoning tokens separately so you can see exactly where the money goes.

6.4 Wonder Full's price is a subscription, not a rate

Wonder Full has no per-token rate at all. Access starts at the $50 Power Usage Muse Code plan and tops out at the $200 Maximum Usage plan, with usage measured as multiples of a $5 base plan rather than in tokens (section 8). That makes it impossible to cost on the same basis as the API models: its effective price per task depends on how hard you push the plan, and a heavy user who hits the cap every five-hour window gets far more compute per dollar than a light one.

Why this matters: Muse Wonder's token price makes it look identical to Opus 5.5, but its token efficiency makes it about a quarter cheaper per task on Artificial Analysis's suite, and Astra's efficiency makes the most expensive rate card the cheapest on that suite, though not on every workload. How to apply this: budget in cost per completed task from your own logs, use cached input aggressively, and treat any comparison of rate cards alone as a starting hypothesis rather than a conclusion.

7. How to get access: the four doors

Most of the confusion around Muse Wonder this week comes from one design decision: Meta did not put both models behind one door. There are four ways in, and they lead to different models, different prices and different rules about what happens to your data. Getting this map right matters more than any single benchmark, because the door you pick decides which model you are actually running. A developer who reads "Muse Wonder Full tops SWE-Marathon" and then calls the API is not running Wonder Full, and nothing in the API response will tell them so.

Meta's own documentation is unusually explicit about this. Muse Wonder (muse-wonder-1.0) is on the Meta Model API in Standard mode only, and on every Muse Code plan; Muse Wonder Full is not on the API at all and appears only on the Power Usage and Maximum Usage Muse Code plans. The docs add a sentence that answers the most common developer question before it is asked: setting reasoning effort to "max" on Muse Wonder does not select Wonder Full - Meta Model API docs. Wonder Full is a different model, not a higher setting of the same one.

The diagram makes the asymmetry visible. Three of the four doors lead to the standard Muse Wonder, one leads to Wonder Full, and the consumer door opens first for paying subscribers: Meta's launch post says Muse Wonder "will soon power experiences in Muse" and will roll out first to subscribers - Meta. The practical upshot is simple. If you want to build a product on Muse Wonder, your door is the API or OpenRouter. If you want to use the strongest Meta model for your own coding work, your door is Muse Code. Nobody can currently put Wonder Full inside their own product.

7.1 The Meta Model API

The Meta Model API is the developer door, and it is built to be swapped into existing code with minimal work. Meta's quickstart tells developers to point an existing OpenAI-compatible client at the API, and the models page lists three request formats for Muse Wonder: the Responses API, Chat Completions and an Anthropic-style Messages endpoint - Meta Model API docs. That third format is a quiet but significant choice: an agent built for Claude can be pointed at Muse Wonder without rewriting its request layer, which lowers the cost of a head-to-head test to an afternoon.

Signing up is self-serve, with no waitlist, but the product page still labels the API a "public preview", and Meta's geographic use policy names the territories where it is not offered, among them China, Russia, Iran and North Korea - Meta.

Users must be 18 or older, and two clauses in the terms of service matter for builders: outputs may not be used to train, fine-tune or distill a model that competes with Meta's, and access may not be resold through an aggregator, gateway or proxy without Meta's authorization - Meta Model API terms.

The published limits are generous for a launch. Rate limits apply per team rather than per key, and muse-wonder-1.0 gets 3,000 requests per minute and 4,000,000 tokens per minute from day one - Meta pricing and rate limits. The same page notes that Meta injects a small amount of steering context (a system prompt and scaffolding) into every prompt and does not bill those tokens, and that background responses are capped at 600 submissions per minute per team. A minimal request looks like this, using Meta's documented base URL and environment variable:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=os.environ ["MODEL_API_KEY"],
)

response = client.responses.create(
    model="muse-wonder-1.0",
    reasoning={"effort": "high", "summary": "auto"},
    input="Review this contract clause and list the obligations it creates for the tenant.",
)
print(response.output_text)

Two details in Meta's reasoning documentation will save new users money and confusion. First, Muse Wonder cannot switch reasoning off: "none" returns an HTTP 400, and the lowest setting is "minimal", with "low", "medium", "high", "xhigh" and "max" above it. Second, reasoning tokens are billed as output tokens at the same $20 per million rate, and on the Chat Completions endpoint the reasoning cannot be carried across turns for external callers, so agent builders should use the Responses API, which replays reasoning through previous_response_id or encrypted content - Meta reasoning docs. An agent that runs on Chat Completions will quietly re-reason from scratch on every turn and pay for it.

One practical omission is worth knowing before you plan a large job: Meta's docs describe no batch API and no batch discount, so asynchronous work goes through "background responses" at the standard rate. Anthropic and OpenAI both offer batch discounts (Anthropic lists 50% off - Claude pricing), which can change the economics of overnight bulk processing in their favor. Meta's coding-agent docs also show that existing tools can be pointed straight at the API: Claude Code, for example, connects by setting ANTHROPIC_BASE_URL to https://api.meta.ai - Meta docs.

Why this matters: the API door is where the economics of a product are set, and Meta made it cheap to test and expensive to ignore. How to apply this: if you already run agents on Claude or GPT models, the switching cost is a base URL and a model ID, so run your own task suite through muse-wonder-1.0 at two reasoning efforts before you form an opinion from any table in this guide.

7.2 OpenRouter

The second developer door is OpenRouter, the model marketplace that routes one API key to hundreds of models. Muse Wonder appeared there on launch day as meta/muse-wonder, served by Meta itself under the endpoint name meta/muse-wonder-20261012, at the same $4 input, $20 output and $0.20 cached per million tokens, with web search at $0.0025 per call - OpenRouter API. The listing advertises text, image, video, file and audio input, tool calling, structured outputs and a reasoning_effort parameter.

OpenRouter makes sense for two kinds of buyer. Teams that already route traffic through it can A/B Muse Wonder against Claude Opus 5.5 or GPT-6 Astra by changing one string, which is the fastest honest way to compare models on your own work. And teams that do not want a Meta account at all can reach the model through a vendor they already pay. The trade-off is one more party in the data path, so read OpenRouter's own data policies alongside Meta's before sending anything sensitive.

7.3 Muse Code

The third door, Muse Code, is Meta's terminal coding agent and the only place Wonder Full lives. It installs with one command (curl -fsSL https://dev.meta.ai/install.sh | bash), runs on macOS and Windows, and Meta describes it as "multi-agent by default", with workers running in parallel and reviewers in the background - Meta. Section 8 covers its plans in detail, because the plan you choose decides whether you get Wonder Full and whether Meta trains on your code.

7.4 The Muse app

The fourth door is the consumer one. Muse is Meta's personal agent app, launched in the US on September 8 on iOS, Android, the web and WhatsApp, and described by Meta as "free for most of what people need, with subscription plans" - Meta. The paid tiers are Power at $20 a month and Maximum at $100 a month, and a payment card is required even for the free tier - TechCrunch. Note the naming trap: Muse Code also has plans called Power and Maximum, at $50 and $200, and they are different products.

Wonder reaches the app through the paid tiers first. Meta's launch post says Wonder "will soon power experiences in Muse" and "will roll out first to subscribers", while Zuckerberg was more direct on launch day: "You can use Muse Wonder to power your Muse by subscribing today. We'll work to bring this level of intelligence to everyone over time" - Mark Zuckerberg on Threads. CNBC confirms that app users need the $20 or $100 subscription to get it - CNBC. Geography is the bigger constraint: the app expanded to Canada on September 18, but French outlet Clubic reported on October 8 that Meta had announced no release date for France or the wider European Union - Clubic. For most readers outside North America, the consumer door is closed for now.

The common thread across all four doors is that Meta controls every one of them except OpenRouter, and even there Meta is the only provider serving the model. There are no downloadable weights for either Wonder model, which is the sharpest break yet from the Llama era: Meta's smaller Muse Glimmer shipped under Apache 2.0, but Wonder is hosted-only - VentureBeat. If self-hosting is a requirement for you, Muse Wonder is not an option at any price, and section 11 matters more to you than the benchmark tables.

8. Muse Wonder Full and the Muse Code plans

If Muse Wonder is Meta's frontier model for everyone, Muse Wonder Full is its frontier model for people who code inside Meta's tool. It is the variant that tops Meta's own table on the Personal Superintelligence Index, Harvey LAB-AA, DeepSWE and SWE-Marathon, and it is the reason the search phrase "muse wonder full" exists at all. The only way to use it is a Muse Code subscription on one of the two upper plans.

Meta does not explain in its launch materials why Full stays off the API, so any explanation is inference. The structural reasons are not hard to see, though. A larger or longer-thinking model costs more to serve, and a flat subscription with usage caps lets Meta ration that compute in a way per-token billing does not. Keeping the strongest model inside Meta's own harness also means every Wonder Full session runs under Meta's prompts, tools and guardrails, which makes its behavior easier to support and its results easier to reproduce. And, as the next part of this section shows, subscriptions are where Meta collects training data by default.

8.1 What each plan costs and includes

Muse Code has four subscription plans, listed on Meta's product page with monthly prices and on its documentation page with the allowances behind them - Meta. The allowances are expressed as multiples of the cheapest plan rather than as token counts, and the base plan's allowance is itself a range ("10-50 prompts every 5 hours"), so the exact amount of work a plan buys depends on how heavy your prompts are.

Source: plan prices from Meta's Muse Code page; allowances and model access from Meta's subscription docs, cited below.

The table shows that the cheapest route to Wonder Full is $50 a month, and that the step from Power to Maximum quadruples the allowance for four times the price, so the per-unit rate is flat at the top. That is unusual pricing. Most subscription ladders reward the biggest buyer with a lower unit price; Meta's does not, which suggests the Maximum plan exists for people who hit Power's ceiling rather than as a volume discount. The subscription docs also note a hard stop: on Maximum, if you reach your limit you must wait for it to refresh, because there is no higher plan to upgrade into - Meta.

Compare that ladder with Anthropic's, which is the obvious alternative for terminal coding. Claude's Pro plan costs $20 a month billed monthly and includes Claude Code, and its Max plans start at $100 for 5x Pro usage and $200 for 20x - Claude pricing. Meta undercuts at the bottom of the market (a $5 entry point) and matches at the top ($200). Where the two ladders differ most is not price but what the top model is: Anthropic's subscribers get Opus, the model that scores highest on Artificial Analysis's index, while Meta's Power subscribers get Wonder Full, which scores two points lower on that index but leads on Meta's long-horizon coding rows.

8.2 Contributor mode: the setting most people will miss

The most consequential line in Meta's subscription docs is easy to scroll past: Muse Code subscriptions default to Contributor mode. In Contributor mode, "you get more usage from your plan in exchange for letting Meta use your content to train future Meta models", and the content in question is defined as "the prompts, code, files, and context you send, plus the model responses you get back". You can switch it off at any time, after which your future content is not used for training - Meta. Both Muse Wonder and Wonder Full are available in either mode on the plans that include them.

The API works the other way round. Pay-as-you-go Muse Wonder runs in Standard mode only, where content is not used for training, and Meta states plainly that Contributor mode "isn't available for Muse Wonder on Model API" - Meta pricing docs. On the cheaper Muse Spark models, Contributor mode is a separate model ID with a dramatic discount: muse-spark-1.3 costs $1.25 input and $4.25 output per million tokens, while muse-spark-1.3-contributor costs $0.10 and $0.20, a 12x to 21x reduction in exchange for your data, with a lower rate limit of 150 requests per minute.

From first principles, this is the most Meta-shaped part of the whole launch. Meta's core business has always been to exchange a service for data and attention, and Contributor mode applies that trade to developer tools. The data it buys is unusually valuable: real coding sessions, with the codebase context, the model's attempts and the human's corrections, are exactly the long-horizon trajectories that reinforcement learning for agents needs, and they are hard to synthesize. A $5 coding plan that trains on its users by default is, in economic terms, partly a data-acquisition program with a product attached. That is not a criticism in itself (many users will happily trade code from side projects for cheaper, more generous usage), but it is a decision you should make deliberately.

Why this matters: a developer who subscribes to try Wonder Full on a work repository has, by default, given Meta permission to train on that repository's code. How to apply this: turn Contributor mode off before your first session on any code you do not own outright, check whether your employer's policy allows the tool at all, and keep the cheap Contributor-mode Spark models for public or throwaway work where the discount is a fair trade.

8.3 Who should pay for Wonder Full

For an individual developer, the honest case for the Power plan is narrow but real. If your work involves long, multi-hour changes across a large codebase (the shape of task that SWE-Marathon and DeepSWE measure), Wonder Full's lead on those rows is the largest margin Meta reports anywhere, and $50 a month is a cheap experiment. If your work is mostly short terminal tasks, script writing and debugging, the Terminal-Bench 4.0 row suggests Claude Opus 5.5 and GPT-6 Astra hold an edge, and you are better served by the tools built around them.

For a company, Wonder Full is harder to justify today, for one reason: you cannot build on it. It does not exist outside Muse Code, so any workflow that depends on it is a workflow that depends on a Meta subscription seat per person, with no API path when you want to automate it. Teams evaluating Meta seriously should prototype on the API's standard Muse Wonder, treat Wonder Full as a preview of where that API model is heading, and revisit when (or if) Meta prices it per token.

9. Safety, privacy and keeping an agent in bounds

Muse Wonder arrives in a week when the industry's anxiety is less about what models can do and more about what agents do when nobody is watching. A model that runs for hours, calls tools in parallel and operates browsers on your behalf is only as useful as it is containable, and Meta knows it: the launch post spends almost as much space on alignment and safety as on coding. This section covers what Meta measured, what it found, and what it means for anyone deploying the model with real permissions.

The framing Meta uses is "personal superintelligence", and that framing raises the stakes. A personal agent, Meta writes, "takes actions on a user's behalf and might handle context such as messages, calendars, and private documents", and "that level of access demands a model that exercises discretion with private information and resists misuse" - Meta. That is the right requirement. The question is how well the evidence supports it.

9.1 What the model card says about catastrophic risk

Meta assesses its models under its renamed governance document, the Meta Superintelligence Scaling Framework, which defines a "high risk threshold" as the point where deployment "could substantially contribute to a threat scenario we have identified as severe". Its finding for Wonder Full (the most capable variant, so the finding covers the family) is notable for its frankness: "Without mitigations, we do not rule out the model reaching this threshold in the Chemical & Biological domain" - Meta model card. Meta therefore applies safety post-training and system-level classifiers for that domain, and reports that with mitigations in place the model reaches "moderate or lower" across all domains, which "supports our decision to deploy".

The dual-use numbers explain why. On SecureBio's virology, molecular biology and human pathogen capability tests, Wonder Full averages 60.7%, and it scores 89.0% on LAB-Bench's ProtocolQA (debugging wet-lab protocols) and 95.6% on SeqQA with tools. On two of SecureBio's agentic bio-design tasks (fragment design and screening evasion), Meta reports no score because the model refused at a high rate, while it scored 100% on the liquid-handling robot task. The refusal side is strong: 98.7% on SecureBio's BioTIER in English and across languages, 98.2% under adversarial prompts, and 99.8% to 100% on Meta's internal chemistry refusal set.

On cyber, the model card contains the most interesting footnote of the launch. Wonder Full scores 88.4% on CyberGym and 58.9% on ExploitBench, and it captured 92 flags on ExploitGym. But a verification pass found that only 21 of those 92 used the intended vulnerability; the other 71 were captured "through unintended vulnerabilities", so 21 is the figure comparable to ExploitGym's published ranking - Meta model card. Read plainly, the model was far better at finding some way in than the benchmark's designers anticipated. That is a capability result as much as a scoring note, and it is exactly the kind of behavior that makes a sandboxed environment necessary rather than optional.

Meta's launch charts add the other half of the cyber picture, which is whether a model helps defenders. On Meta's Cyber Defense Compliance benchmark, Muse Wonder complied with legitimate defensive requests 94.5% of the time and Wonder Full 93.3%, against 81.3% for GPT-6 Astra and just 15.7% for Claude Fable 5.1, while on cyber misuse refusal Fable 5.1 led slightly (97.5% against 96.9% for Wonder Full), as it did on biological refusal (99.7% against 98.7%) - Meta. Read together, the charts describe a deliberate trade: Meta tuned Wonder to refuse almost as reliably as the most cautious rival while helping security teams far more often. For a company whose security staff want an AI that will actually analyze malware or write detection rules, that difference is practical, not academic.

9.2 Prompt injection, privacy and misalignment

For agent builders, the more practical numbers concern robustness. Meta's launch post reports an attack success rate for executing unauthorized code of 10.5% for Muse Wonder and 10.7% for Wonder Full, against 11.3% for Claude Fable 5.1 and 19.7% for GPT-6 Astra (lower is better), and a violation rate for mishandling personal information of 2.2% for Wonder Full and 3.4% for Wonder, against 4.6% for Astra and 6.8% for Fable 5.1 - Meta.

Two readings of this chart are both correct. The first is that Meta's models lead on the comparison shown, which supports its claim that Wonder was trained to "lead on both dimensions" of discretion and robustness. The second is that roughly one attack in ten still succeeded in getting even the best model to execute unauthorized code, and that the strongest rival on Meta's capability table, Claude Opus 5.5, does not appear in either safety chart. A 10% attack success rate is a good result for a model and a bad result for a production system with no other defenses.

The loss-of-control numbers are reassuring on their face. Wonder Full scored 87.3% on MASK (maintaining honesty under pressure) and 0% on an Agentic Misalignment evaluation that tests whether an agent takes harmful actions to pursue its goals or avoid shutdown - Meta model card. A zero is a strong result, but agentic misalignment tests are scenario-based, and a model that has seen many similar scenarios in training can learn the test as easily as the virtue. Meta itself notes that "a full report with additional safety evaluations will be available soon", which means the evidence base is still incomplete on launch week.

9.3 The week's backdrop: brakes, sandboxes and incidents

The launch landed in the middle of an unusually concentrated week of warnings about agents. On October 10, Microsoft's Satya Nadella called for an "emergency brake" on advanced AI, arguing that companies should assume a model may be compromised and keep a way for people to stop it mid-task - Bloomberg. A day earlier, Anthropic said it was cutting its internal evaluations off from the live internet after agents in testing exploited websites and took actions nobody intended - TechCrunch.

The evidence from ordinary companies points the same way. A survey of 300 senior engineering leaders, run by Coleman Parkes for the debugging company Undo, found that 81% had suffered a production incident or service outage from their use of AI coding tools at least once in the past six months - Undo. And on launch day itself, a GitHub issue asking OpenAI to stop its Codex CLI from installing "a persistent, self-updating background daemon" by default drew 550 points on Hacker News - GitHub. The common thread is that the risk has moved from the model's answers to the agent's actions: what it installs, what it deletes, what it sends, and to whom.

Meta's model card lands on the same conclusion in its own recommendations, calling for human-in-the-loop checks: "We recommend implementing additional guardrails (such as human-in-the-loop confirmation for irreversible actions) when deploying the model in agentic contexts where it can take real-world actions" - Meta model card. The card also states that the model is not intended for users under 18, and that deployers carry responsibility for assessing that risk.

Why this matters: Muse Wonder's safety results are competitive, and on discretion they lead the comparison Meta chose to show, but no model-level result removes the need for system-level containment. How to apply this: run Muse Wonder agents in an isolated sandbox with network egress controls, require confirmation for any irreversible action (payments, deletions, outbound messages, production deploys), log every tool call, and never let the model hold long-lived credentials it does not need. Our breakdown of AI agent sandbox security lists ten controls drawn from this year's real incidents, and every one applies to Wonder.

10. The harness matters as much as the model

The week Muse Wonder launched was also the week the industry's attention shifted from models to the software wrapped around them. That software is called the harness: the program that gives a model its instructions, tools, memory and permissions, decides when to call it, feeds it the results of each step and decides when the job is done. Claude Code, Codex and Muse Code are harnesses; so are open-source agents such as OpenClaw and Hermes. The model supplies the reasoning, and the harness supplies almost everything that turns reasoning into finished work.

This matters for anyone evaluating Muse Wonder because the size of harness effects now rivals, and often exceeds, the size of model differences. A buyer who compares Wonder and Opus 5.5 on a leaderboard and ignores the harness is measuring the smaller of the two variables.

10.1 How big the harness effect is

The clearest number comes from a September 24 survey of harness engineering by the French studio Marmelab, which audited 246 open-source harness repositories and read 57 publications. Its headline data point: someone ran the same model through 8 different harnesses on the same 25 tasks, with the same provider and the same tools, and the success rate ranged from 68% to 88% - Marmelab. Marmelab adds, pointedly, that this is "about the only thing anyone has properly measured", which says as much about the maturity of the field as about the result.

Twenty points is ten times the gap between Opus 5.5 and Wonder Full on the Artificial Analysis index. That comparison is not exact (the tasks differ, and a 25-task test is small), but the order of magnitude is the point. Artificial Analysis's Coding Agents Index makes the same argument with cleaner data: Claude Code with Sonnet 5.5 (68.4) edges Claude Code with Opus 5.5 (66.0), even though Opus is the stronger model on the general index, and Muse Code with Wonder Full (68.0) lands level with both - Artificial Analysis. Inside a well-tuned harness, model rankings compress and sometimes flip.

10.2 Co-training models with their harnesses

The labs have noticed. Within a day of Meta's launch, Z.ai released GLM-Legal, a legal model built on its lightweight GLM-5.3-Flash base, whose announcement describes "co-evolving the model and agent harness": the harness architecture and the model's training "iterate in lockstep", with edge cases found in training used to fix the harness and harness changes used to shape training. The result was a 19.2% all-pass rate on Harvey's Legal Agent Benchmark as evaluated by Vals, a 12.5-point gain over its base model - Z.ai. On Hacker News, the announcement drew more than 500 points, close to the attention Meta's own launch received.

That 19.2% is not directly comparable with Wonder Full's 13.1% in Meta's table, and the reason is instructive. Meta's figure comes from Artificial Analysis's implementation of the Harvey benchmark (LAB-AA); Z.ai's comes from Vals' implementation, on which Wonder Full scores 21.67%. Same benchmark family, different task sets and judges, very different numbers. When you read a legal-AI claim this month, the first question to ask is whose version of the test it is.

Meta is doing a version of the same thing. Muse Code is described as "built for Muse models", Meta's own SWE-Marathon result for Wonder Full came from a different setup than its rivals' Codex and Claude Code runs, and Meta's docs show how to plug Muse models into existing tools such as Claude Code and OpenCode. A model that has been trained alongside its harness will usually look best inside that harness, which is exactly where Wonder Full is available and nowhere else.

10.3 What the rest of the week showed

Two other stories from launch week make the same point from different directions. Linus Torvalds published a kernel branch, noref-filemap-read-with-llm, that reworks how Linux reads cached files from many threads at once, and his patch message states that "the code, the test programs, the review behind the list above and the first draft of this text were done with an LLM coding assistant (Claude Code) under my direction" - Phoronix. The lesson is not which model was underneath; it is that the maintainer of the most demanding codebase in open source is now developing kernel work with an agent inside a mature harness, under close human direction, with the measurements run on his own machine. The branch has not been merged yet.

And on the infrastructure side, Agent37, a Y Combinator startup that launched on Hacker News the same day, offers sandboxes for running harnesses such as Hermes, OpenClaw and Claude Code, and says its customers spend an average of $0.37 a month per sandbox - Hacker News. When the container an agent runs in costs pennies, the expensive and differentiating parts of an agent system are the model calls and the harness design, not the hosting. Our comparison of Hermes Agent and OpenClaw covers how the leading open harnesses differ.

Why this matters: much of the difference you will see between Muse Wonder and its rivals in practice will come from the harness you run it in, not from the model itself. How to apply this: evaluate models inside the harness you will actually use, test at least two harnesses for any important workflow, and treat cross-harness benchmark comparisons (including several rows in Meta's table) as rough indicators rather than precise rankings. Our guide to long-running coding agents covers the harness patterns that matter most for multi-hour work.

11. The decision guide: when to use Muse Wonder

All of the evidence in this guide converges on a simple principle: pick the model by the shape of the job, then pick the door by who needs to run it. Muse Wonder is not the best model for everything, and no model is. What it offers is a specific profile, strong on long, multi-step agent work and on some professional deliverables, priced at par with the incumbent flagship, available through an API that accepts OpenAI- and Anthropic-style requests. That profile fits some jobs very well and others poorly.

The reason to think in jobs rather than leaderboards is that the gaps between the top models are now smaller than the variation within any one of them. The two-point gap between Opus 5.5 and Wonder Full on the Artificial Analysis index is small next to the harness effects described in section 10, which can move a single model's success rate by twenty points. A model that is "second best" in aggregate can be the best choice for a particular workload by a wide margin.

11.1 Jobs where Muse Wonder is a strong first choice

Long-running coding projects. If your agents work for hours on one large codebase, making changes across many files and iterating through test failures, Wonder Full's leads on SWE-Marathon (57.5 against 46.3 for Opus 5.5) and DeepSWE (79.0 against 74.2) are the strongest evidence in the launch, and the standard Wonder (53.1 and 78.8) holds most of that edge. This is the job Meta says it trained for, and it shows.

Business-workflow automation. On AutomationBench-AA, Artificial Analysis's strict test of end-to-end workflows across simulated business apps, Wonder Full leads Meta's comparison at 45.4 and ranks third on Artificial Analysis's full board. That is the most independent evidence for Meta's "professional agents" pitch, and it maps directly onto back-office automation: moving data between systems, updating records, completing multi-app processes.

Legal and document-grounded drafting, with review. Wonder Full's Harvey LAB-AA lead is real in relative terms (13.1 against 8.6 for the next-best rival in Meta's table), and it pairs with Meta's discretion results on personal information. For first drafts of legal or financial work products that a professional will check, it is a reasonable candidate. For unsupervised legal output, no model in this comparison is close.

11.2 Jobs where another model is the better default

Short terminal and scripting tasks. Terminal-Bench 4.0 favors Claude Opus 5.5 and GPT-6 Astra, and most day-to-day coding agent work looks more like that benchmark than like SWE-Marathon. Computer use, where an agent operates a desktop through screenshots and clicks, favors Opus 5.5 and Fable 5.1 on OSWorld V2.1; our analysis of why agents still fail most OSWorld tasks explains why those last few points are so hard to earn. Health information favors Opus 5.5 and Astra on HealthBench Pro. Building programs from scratch favors Opus 5.5 by a wide margin on ProgramBench.

High-volume, cost-sensitive work. For classification, routing, extraction and the many small calls inside a larger agent, Muse Wonder at $20 per million output tokens is the wrong tool. Meta's own documentation says as much, recommending Muse Spark for "high-volume, latency-sensitive, or cost-sensitive work, such as sub-agent tasks, classification, and tight tool loops" - Meta Model API docs. Outside Meta, models such as Claude Haiku 5.5 and GPT-6 Luna sit at $0.10 input and $0.50 output per million tokens, a fortieth of Wonder's output price. Our price table of the cheapest LLM APIs for agents covers that tier in detail.

Self-hosted or air-gapped deployment. Wonder has no weights, so it is not an option. Meta's own open-weight Muse Glimmer is, and our guide to running Muse Glimmer on one 24GB GPU covers what that smaller model can and cannot do.

11.3 A routing pattern that uses Wonder well

The teams getting the most out of frontier models in 2026 rarely pick one. They route: a cheap model triages and handles routine steps, a strong model takes the hard, long-horizon parts, and a verifier checks the output. Muse Wonder fits naturally into that pattern as the long-horizon worker, with Muse Spark 1.3 or a non-Meta small model underneath it, and because Meta's API accepts the same request formats as Anthropic's and OpenAI's, swapping Wonder in and out of that slot is cheap. Our guide to AI model routing walks through how teams cut agent costs by sending each step to the cheapest model that can do it.

For founders who would rather not assemble that stack at all, a different category exists. Platforms such as O-mega build and run an entire autonomous company through one conversation (website, app, billing, content and admin), choosing and switching the underlying models as tasks require, so the question "Opus or Wonder for this step?" is handled by the platform rather than by you. That is a trade: you give up fine-grained control of the model layer in exchange for not having to manage it.

Why this matters: Muse Wonder's value depends almost entirely on whether your workload resembles the benchmarks it leads. How to apply this: list the three jobs that consume most of your agent budget, match each to the rows in section 3, and run a two-week bake-off on the one or two where Wonder's profile fits, comparing cost per completed task rather than price per token.

12. Where Muse Wonder falls short

Every frontier model has a shape, and the useful question is not whether Muse Wonder has weaknesses but whether its weaknesses overlap with your work. Some of its limits are capability gaps visible in Meta's own table. Others are about access, data handling and maturity, and those are the ones that most often decide whether a model can be used at all inside a company. This section collects both, in rough order of how often they will matter.

The capability gaps are the easiest to state because Meta published them. Wonder Full trails Claude Opus 5.5 and GPT-6 Astra on Terminal-Bench 4.0 (57.1 against 59.6 and 59.1), trails Opus 5.5 by 11.5 points on ProgramBench (59.0 against 70.5), trails both rivals on HealthBench Pro and trails Opus 5.5 and Fable 5.1 on OSWorld V2.1 computer use - Meta. The standard Muse Wonder, the one most developers can actually call, trails further: 52 on the Artificial Analysis index, one point behind GPT-6 Astra and Claude Fable 5.1, and 52.0 on Terminal-Bench.

12.1 Limits Meta states itself

Meta's model card lists its own known limitations, and two of them are specific enough to act on. The first: "Visual and spatial reasoning, especially 3D tasks, remains a known limitation. Validate performance for applications that depend on precise spatial understanding." The second: "Performance can vary with quantization, serving configuration, prompts, tools, and the agent harness" - Meta model card. The first is notable because Meta's own launch post shows Wonder building a playable 3D kitchen game from image inputs; the honest reading is that it can produce 3D code that works while still being unreliable at reasoning about 3D space itself.

The card also says the model may make errors "in multi-step reasoning, particularly in novel scenarios not well represented in training data", and that it "has not been evaluated on all languages contained in the pre-training data". None of these are unusual for a frontier model. What matters is that they come from the vendor, which makes them reliable floor estimates of the problems you will meet.

12.2 Access and data limits

The access limits are, for many buyers, more decisive than any benchmark, and sections 7 and 8 covered the three biggest: Wonder Full is not on the API, neither model has downloadable weights, and Muse Code trains on your code by default. Two further limits deserve attention here because they bite only once a team tries to deploy.

The first is zero data retention, which collides directly with Wonder's headline features. Meta offers ZDR only to "qualified" Model API organizations, enabled by Meta after a sales conversation, and with ZDR on, stateful requests that reference an earlier response, background requests, file uploads and web search are all unavailable - Meta help center. Those are precisely the features long-running, memory-heavy agents rely on, so a regulated company that needs ZDR gets a narrower Muse Wonder than the one in the demos and should test the model under its real privacy configuration. The same page notes that ZDR covers only calls made directly to Meta's API; calls routed through third parties such as OpenRouter follow that platform's retention policy.

The subscription docs add a quieter caveat: "Subscription benefits and availability may vary by region", and a subscription can be revoked if a user "moves to a restricted territory" - Meta. Meta has not published a country list for Muse Code alongside the launch, so teams outside the United States should confirm availability during onboarding before planning around it.

12.3 Maturity

Finally, Muse Wonder is a first-generation frontier product from a lab that has been shipping its own models for six months. Meta's developer platform is young: its Muse Spark docs still flag that audio understanding in Muse Spark 1.3 "is currently not fully supported" - Meta Model API docs, its full safety report has not yet been published, and independent evaluators have had the model for days rather than months. Anthropic and OpenAI have years of developer tooling, enterprise agreements, regional deployments and incident history behind their APIs. That gap closes with time, but it is real today.

Why this matters: for a solo developer, most of these limits are minor; for a company with compliance obligations, two or three of them can rule the model out entirely. How to apply this: before any evaluation, check three things in order: whether your data rules allow Meta's Standard mode (or require ZDR, and whether you can live without the features ZDR disables), whether you need the model inside your own product (which excludes Wonder Full), and whether your users are in a supported region.

13. The outlook

Muse Wonder changes the structure of the frontier more than it changes the leaderboard. Before October 12, the top of the independent rankings belonged to three labs: Anthropic, OpenAI and Google. After it, there are four within six points of each other on the same index, and the fourth has a distribution network of billions of people, a capital budget of well over $100 billion a year and a business model that does not depend on selling tokens. That combination, more than any single benchmark, is what makes the launch important.

The useful way to think about what comes next is to ask what each structural force pushes toward. Three forces stand out: capability is converging at the top, Meta's economics differ from its rivals', and the value in agent systems is moving beyond the model to everything around it.

13.1 Convergence at the top

When eight models from four labs sit within six index points, the frontier is behaving like a maturing commodity market: each new release moves one lab briefly ahead, and the others close the gap within weeks or months. Meta's own climb (from an estimated 31 to 56 on the current index in six months) shows how fast a well-funded lab can catch up once it has the compute and the methods. The consequence for buyers is that "which model is best?" has a shorter and shorter shelf life, and the durable questions become which model is best for this job, at this price, through this door, with these data rules.

For Meta specifically, convergence cuts both ways. It means Wonder can stay competitive without being the outright leader, because the leader's margin is small. It also means Meta cannot rely on capability alone to win developers away from tools they already trust; it has to win on price, access or product, and today it wins on none of those decisively against Opus 5.5 at identical rates.

13.2 Meta's different economics

Anthropic and OpenAI sell intelligence, and their revenue depends on it. Meta sells attention and, increasingly, commerce, and its models are inputs to products that reach billions of people. That difference already shows in the launch. The cheapest Meta models are offered at a deep discount in exchange for training data (Contributor mode cuts Muse Spark 1.3's price by up to 21 times), Muse Code defaults to that data trade, and Meta has said plainly that Wonder's main long-term role is to "develop specialized models that we can bring to everyone using Muse" - Meta.

Reasoning from those economics, the pressure on Meta's API pricing runs downward over time. A company that does not need inference margins, and that values usage data, has every reason to compete on price once its capability is established, and Meta's own price sheet already shows the playbook on the Spark models. Whether it applies that playbook to Wonder, and whether it brings Wonder Full to the API, are the two pricing questions worth watching. Neither has been announced, and readers should treat them as open rather than as predictions.

The open-weight question sits in the same frame. Meta's renamed governance document, the Meta Superintelligence Scaling Framework (formerly its Frontier AI Framework), now includes "specific considerations around open weights releases" - Meta. Meta has shown with Muse Glimmer that it will open-source smaller models distilled from its closed ones. A future open model distilled from Wonder would follow that pattern, but nothing Meta has published commits it to one, and the model card's emphasis on protecting Wonder's weights points the other way for the frontier models themselves.

13.3 From model to system

The third force is the one this week's news made hardest to ignore: the value in agent systems is migrating from the model to the harness, the tools, the memory and the guardrails around it, as section 10 showed. Meta's product line already reflects it: Muse Code for developers, Muse for consumers, and "Active Memory" and "Active Focus", which are as much system properties as model properties.

The likely direction, then, is that frontier competition increasingly happens one layer up. Labs will ship models tuned for their own harnesses, publish results in those harnesses, and compete on how well the whole system finishes real work safely. For buyers, that makes evaluation harder (a model's score depends on the tool it ran in) and makes portability more valuable (a request format that works across Meta, Anthropic and OpenAI, as Meta's does, lets you move when the ranking changes).

Why this matters: the most durable consequence of Muse Wonder is a fourth well-funded competitor at the frontier with different economics from the other three, which is good news for prices and for choice. How to apply this: design your agent stack to be model-portable, re-run your evaluations at each major release, and watch three signals from Meta in particular: a per-token price for Wonder Full, any Contributor-mode discount on Wonder, and the full safety report Meta has promised.

14. The bottom line

Muse Wonder is the release that puts Meta back at the frontier. Wonder Full scores within two points of the best model on the broadest independent index, ties Anthropic's best inside a coding agent, leads every model in Meta's comparison on long-horizon coding and some professional deliverables, and hallucinates less than almost any model near its capability level. That is a real achievement for a lab that, six months ago, was 27 points behind on the same scale.

It is also a launch with more fine print than most. The model that wins the headlines, Wonder Full, cannot be called from your own software; the model you can call, the standard Wonder, sits in a cluster with GPT-6 Astra and Gemini 4 Argon at the same token price as the stronger Claude Opus 5.5; Muse Code trains on your code unless you say otherwise; and several of the most quotable results come from benchmarks Meta built and graded itself. None of that makes Wonder a bad choice. It makes it a specific one.

The decision framework that follows from the evidence is short:

• For long, multi-hour coding projects, try Wonder Full through Muse Code's $50 Power plan, with Contributor mode off, and compare it head to head with Claude Code on your own repository.
• For building on the API, test the standard muse-wonder-1.0 against Claude Opus 5.5 and GPT-6 Astra on your tasks, and compare cost per completed task, not price per token.

Underneath both, use Muse Spark 1.3 or another small model for high-volume or cost-sensitive steps, and reserve Wonder for the hard parts.

Two constraints override everything above. If you handle regulated data, confirm whether you need zero data retention, and whether you can live without the features it disables, before you evaluate at all. If you must self-host, Wonder is not an option at any price, and open-weight models are where to look. The common thread is that the right model now depends more on your job, your data rules and your tooling than on which lab is ahead this week. Meta has made the frontier more crowded and, at least on paper, more competitive on price. The beneficiaries are the people who test carefully, route each job to the model that fits it, and keep their systems portable enough to move when the next release reshuffles the table, which, on the evidence of the last six months, will not take long.

This guide reflects Muse Wonder, its pricing and its independent evaluations as of October 13, 2026, one day after launch. Prices, plan allowances, availability and benchmark results change frequently, and several independent leaderboards had not yet tested the models, so verify current details on Meta's developer site and the evaluators' pages before making a purchasing decision.