Grok 4.5: Benchmarks, Pricing & Performance (2026)

The complete, source-checked breakdown of SpaceXAI's Grok 4.5: what the benchmarks actually say, what it really costs, and why the most interesting number is not on the leaderboard.

Grok 4.5 does not top the coding leaderboards it launched with. It loses two of the four benchmarks xAI chose to publish, an independent evaluator ranks it fourth on raw intelligence, and the model its own creator compared it to (Claude Opus 4.7) is already a version behind. And yet Grok 4.5 might be the most consequential model release of July 2026, because the story is not the score. It is the machine that produced it.

On July 8, 2026, a company that did not exist under that name six months earlier shipped a frontier model - TechCrunch. Grok 4.5 is the first model from SpaceXAI, the entity created when SpaceX absorbed xAI, and it is the first model co-trained on data from Cursor, the AI coding tool that same company agreed to buy for $60 billion. One owner now controls the compute, the model, and the surface where millions of developers write code. Grok 4.5 is the first product of that vertical stack, and it is priced to move: $2 per million input tokens and $6 per million output tokens, roughly a third the cost of its closest rivals.

Here is the problem with almost every take you will read on Grok 4.5. Half of them repeat xAI's own benchmark chart as if it were independently verified (it is not), and the other half dismiss the model as mid-pack and move on (which misses why it exists at all). Neither helps you decide whether to route production traffic to it, pay for it inside Cursor, or ignore it. The honest picture is more interesting than either: a model that is not the smartest, is close to the cheapest, and is attached to the largest data-and-distribution flywheel in the industry.

This guide breaks Grok 4.5 down from first principles. It covers exactly what the model is, the corporate story that produced it, every published benchmark and what the numbers hide, the token-efficiency claim that is the real economic pitch, the complete pricing sheet across API and consumer tiers, how it performs in practice, where it wins and where it fails, how it stacks against GPT-5.6, Opus 4.8, Fable 5, Gemini, and DeepSeek, the safety and trust questions, and what the roadmap says about Grok 5. We start high level, then go deep.

Contents

1. What Grok 4.5 actually is (the 60-second version)
2. From xAI to SpaceXAI: the corporate story behind the model
3. The Cursor acquisition and the data flywheel
4. The full benchmark breakdown (and what the chart hides)
5. The token-efficiency claim: the benchmark table is not the cost table
6. Pricing in full: API, consumer tiers, and hidden surcharges
7. Performance and architecture in practice
8. Where Grok 4.5 wins and where it fails
9. How it stacks up against the field
10. Safety, controversy, and the trust question
11. The roadmap: Grok 5 and what comes next
12. What this means for you: a decision framework

1. What Grok 4.5 actually is (the 60-second version)

Before the corporate drama and the benchmark arguments, it helps to state plainly what this model is. Grok 4.5 is a reasoning-first, coding-and-agentic model built on a 1.5-trillion-parameter mixture-of-experts foundation that xAI internally calls V9 - CryptoBriefing. It is positioned not as a general chatbot but as a workhorse for software engineering, agentic tool use, and knowledge work, which is why its launch materials led with coding benchmarks rather than the usual math and trivia leaderboards. Elon Musk described it as an "Opus-class model, but faster, more token-efficient and lower cost" - PYMNTS.

The specifications are straightforward and, unusually for xAI, mostly confirmed by primary sources. Grok 4.5 has a 500,000-token context window, serves at roughly 80 tokens per second, and costs $2 per million input tokens, $0.50 per million cached input, and $6 per million output - SpaceXAI developer docs. It accepts text and image input, produces text output only, and runs a reasoning process at configurable effort levels (low, medium, high) with high as the non-disableable default. It is available from day one inside Cursor, through the SpaceXAI API console, and via a new terminal coding agent called Grok Build.

What matters most is the framing. Musk did not claim Grok 4.5 was the smartest model in the world, and the benchmarks back that restraint. The pitch is value, not supremacy: a model good enough to sit in the frontier conversation, priced and engineered to cost far less per completed task. That single strategic choice explains almost everything about the release, from the benchmarks xAI chose to publish to the acquisition that fed the model its training data. We covered the broader shape of this market in our May 2026 benchmark and pricing roundup, and Grok 4.5 slots into it as the clearest example yet of a lab competing on economics rather than raw capability.

If you remember only three facts about this model, make them these. It is not the smartest model available, ranking fourth on the leading independent intelligence benchmark behind Fable 5, GPT-5.5, and Opus 4.8 - Artificial Analysis. It is close to the cheapest capable model, at a third the token price of its rivals and using far fewer tokens per task. And it is attached to the most powerful data-and-distribution engine in the industry, sitting inside a coding tool used by millions and owned by the same company that trains it. Every practical decision about whether to use Grok 4.5 flows from holding those three facts at once, rather than fixating on any one of them. The rest of this guide is the evidence behind each, and the judgment calls they add up to.

2. From xAI to SpaceXAI: the corporate story behind the model

You cannot understand Grok 4.5 without understanding who made it, because the maker changed identity three times in five months. The structural question is not "did xAI ship a good model," it is "what does it mean when a rocket company, a social network, a supercomputer, and a coding tool all report to the same owner." That consolidation is the actual product. The model is a byproduct.

The sequence is documented and confirmed by major outlets. In January 2026, xAI closed a $20 billion Series E at a $230 billion valuation, with Nvidia, Fidelity, the Qatar Investment Authority, and Abu Dhabi's MGX participating - CNBC. Weeks later, on February 2, 2026, SpaceX absorbed xAI in an all-stock deal valuing the combined company at $1.25 trillion (SpaceX at $1 trillion, xAI at $250 billion), described at the time as the largest merger ever recorded - CNBC. xAI became a wholly owned SpaceX subsidiary, and Grok became, in effect, SpaceX's AI division.

To grasp why the identity kept changing, follow the valuations, because each step was a repricing of the same underlying bet. xAI was worth about $80 billion when it absorbed X in March 2025, roughly $200 billion at a September 2025 raise, and $230 billion by its January 2026 Series E, before the SpaceX merger folded it into a $1.25 trillion whole - Sacra. That trajectory is not the story of a chatbot company growing up. It is the story of an owner assembling the pieces of a fully integrated AI enterprise (compute, model, distribution, and now the developer surface) and repricing the combination upward at each step. Michael Nicolls, a former vice president of SpaceX's Starlink, was installed as president of the AI unit in April 2026, a signal that the group was being run as core SpaceX infrastructure rather than a semi-independent research lab.

The consolidation then accelerated. In May 2026, Musk announced xAI would cease to exist as a separate company, folded entirely into SpaceX with its AI products rebranded under a new banner - Yahoo Finance. On June 12, 2026, SpaceX went public on the Nasdaq under the ticker SPCX, raising roughly $75 billion (up to $85.7 billion including the greenshoe) at a valuation near $1.77 trillion, the largest IPO on record - CNBC. Finally, on July 6, 2026, the AI unit officially rebranded as SpaceXAI, changing its handle to @SpaceXAI and unveiling a logo that folds the old xAI mark into the SpaceX emblem. Two days later, Grok 4.5 arrived as the first model to carry the new name. We traced how this fortune was assembled in our guide to the SpaceX IPO and how Musk became a trillionaire, and the raw scale of the AI business it now houses is worth pausing on.

The numbers from SpaceX's S-1 filing are the first audited window into this business, and they cut both ways. The AI-and-X segment posted $3.2 billion in 2025 revenue against a $6.4 billion operating loss and $12.7 billion in capital expenditure - PitchBook. It disclosed 550 million combined monthly active users across X and Grok, of which roughly 117 million actively use Grok features, up from about 35 million in December 2025 - Forbes. This is the paradox that Grok 4.5 is built to address: enormous reach, enormous losses, and a burning need to convert cheap intelligence into revenue faster than the compute bill grows. A model that costs a fraction of its rivals to run, and that plugs directly into a coding tool used by millions, is exactly what that balance sheet demands.

3. The Cursor acquisition and the data flywheel

If the SpaceXAI merger explains the compute and the capital, the Cursor acquisition explains the data. And data, not parameters, is where Grok 4.5's real advantage is supposed to come from. To see why, start from a first principle about how frontier models improve: past a certain scale, raw web text yields diminishing returns, and the scarce input becomes high-quality demonstrations of expert work. Nowhere are those demonstrations more abundant, or more precisely labeled by outcome, than in an AI coding tool where millions of engineers accept or reject the model's suggestions thousands of times a day.

That is what SpaceX bought. On June 16, 2026, SpaceX agreed to acquire Anysphere, the maker of Cursor, for roughly $60 billion in an all-stock deal, structured as a reverse triangular merger and expected to close in the third quarter of 2026 - TechCrunch. It is the largest acquisition of a venture-backed startup ever recorded, valuing Cursor at around fifteen times revenue after the tool reached an estimated $2 billion in annual recurring revenue in early 2026. We broke down the full logic of that deal in our dedicated guide to why SpaceX bought Cursor for $60 billion, and it argued the purchase was never about owning a text editor. It was about owning the surface where developers interact with AI, and the data that surface generates.

The scale of that surface is what makes it worth $60 billion. Cursor reached roughly 7 million monthly users and more than 50,000 engineering teams, with adoption spanning around two-thirds of the Fortune 500, after growing from about $100 million in annual recurring revenue in early 2025 to an estimated $2 billion a year later and a reported $4 billion annualized run-rate by May 2026 - Yahoo Finance. SpaceX did not commit to the full purchase in one move. It first secured an option in April 2026 to either buy Anysphere for $60 billion or walk away for around $10 billion in combined breakup and joint-work fees, then converted that option into a definitive agreement in June once the value of the data pipeline was clear. Owning Cursor also solves Cursor's own bottleneck: the tool gains direct access to the Colossus supercluster, ending the compute constraints that had capped its ambitions as an independent company.

The mechanics of how that data reached Grok 4.5 are more modest than the marketing implies, and the honesty here matters. According to Cursor's own co-launch post, Grok 4.5 was trained on "trillions of tokens of Cursor data" capturing real developer sessions, but that data was added as supplemental post-training, not baked in from the start - Cursor blog. The V9 base finished pre-training around May 25, 2026, before the acquisition even closed, so an xAI engineer conceded the Cursor data was folded in late and that this is "not quite as good as having it in initial training" - Digital Applied. The next model, reportedly a roughly 2-trillion-parameter successor, is intended to include Cursor data from the beginning of pre-training.

Step back and the strategic logic is worth stating from first principles, because it is the real reason this deal happened. Every frontier lab can buy compute and hire researchers, so neither is a durable moat. What cannot be easily bought is a continuous stream of expert demonstrations labeled by real outcomes, and that is exactly what an owned coding tool produces: every accepted edit is a positive example, every rejected one a negative, every debugging session a trace of how a professional actually solves a problem. By owning Cursor, SpaceXAI converts millions of developers' daily work into a proprietary, self-refreshing training set that competitors cannot replicate without a comparable surface of their own. That is why the acquisition is better understood as buying a data pipeline than buying a product, and it is why the flywheel, not the benchmark score, is the thing rivals should actually fear. The near-term model is only "Opus-class." The long-term threat is a training loop no one else can run.

There was also an embarrassing wrinkle that tells you how raw this pipeline still is. Cursor disclosed that an earlier snapshot of the Cursor codebase was accidentally included in Grok 4.5's training data, giving the model an unfair advantage on CursorBench, so Cursor pulled that benchmark from the comparison while it rebuilds the test suite - Cursor blog. It is a small contamination, quickly disclosed, but it is also a preview of the central tension in this whole flywheel: when the company that trains the model also owns the tool that grades it, every benchmark becomes a question of trust. That tension runs straight through the numbers we are about to examine.

4. The full benchmark breakdown (and what the chart hides)

Grok 4.5 launched with a single benchmark chart, and that chart is the most-shared and least-understood artifact of the release. The image the user of any AI tool is most likely to have seen shows four rows and five columns, with Grok 4.5 highlighted and appearing to hold its own against the best models in the world. Read carefully, with the missing context restored, it tells a more measured story: Grok 4.5 is a genuinely competitive coding model that is nonetheless beaten, on most of these very benchmarks, by at least one rival.

It helps to know what these tests actually measure, because the names are opaque and the differences matter. SWE-Bench Pro presents a model with real GitHub issues from real open-source projects and scores whether its code patch actually fixes the bug and passes the project's tests, which makes it the closest thing to a "can this model do a developer's job" exam. Terminal-Bench measures whether a model can operate a command line to complete multi-step tasks, the skill an autonomous coding agent lives or dies on. SWE-Bench Multilingual runs the same fix-the-bug challenge across many programming languages rather than just Python, and DeepSWE is a harder agentic variant where the model must plan and iterate across a codebase. The reason a model can win one and lose another is that each stresses a different muscle: raw code correctness, tool operation, language breadth, or long-horizon planning. A single number never captures all four, which is exactly why launch charts that show four rows deserve more scrutiny than a single headline score.

Here is the full set of published numbers, restored to the four headline benchmarks plus the two that circulated alongside them. Note the effort levels, which are not cosmetic: Grok 4.5 was measured at effort high, while Opus 4.8 and Fable 5 ran at max and GPT-5.5 at xhigh - Agentpedia.

The single most important correction to the popular reading is this: against Opus 4.8, Grok 4.5's real record on the four headline benchmarks is two wins and two losses, not the rosier picture the highlighted column implies - Roo. It beats Opus on Terminal-Bench 2.1 and DeepSWE 1.0, and loses to it on SWE-Bench Multilingual and SWE-Bench Pro. When a third party (DataCurve) re-ran DeepSWE 1.1 on a neutral harness rather than each vendor's own, Grok 4.5 dropped to fourth of five, behind Fable 5, GPT-5.5, and Opus 4.8 - The Decoder. And Fable 5, Anthropic's most capable model, quietly tops nearly every row it appears in.

Two structural caveats deserve to be stated plainly, because they change how much weight any of these numbers can bear. First, every figure in that chart is self-reported by xAI, including the scores it attributes to competitors, each run on its own harness rather than verified by Anthropic, OpenAI, or a neutral lab - The Decoder. Second, xAI published only coding and agentic benchmarks, and no knowledge or reasoning tests (GPQA, AIME, Humanity's Last Exam, MMLU-Pro) for Grok 4.5, unlike the broad suites it published for Grok 4 - TechCrunch. Any GPQA or AIME numbers you see circulating for "Grok 4.5" almost certainly belong to Grok 4. This is a narrow, carefully chosen window onto the model, and it is the reader's job to remember that a window is not the room.

The effort-level asymmetry is the subtlest distortion, and it works in Grok's favor in a way most readers miss. Reasoning models can be dialed to spend more or less compute per problem, and higher effort usually buys higher scores at the cost of more tokens. Grok 4.5 was benchmarked at high, one notch below the max setting used for Opus 4.8 and Fable 5. That framing lets xAI show a leaner, cheaper configuration of its own model against the most expensive configuration of its rivals, which flatters both its scores relative to cost and its token-efficiency story. It is not dishonest, since the settings are disclosed, but it is a choice that shapes the impression, and a like-for-like comparison at matched effort would likely narrow Grok's wins and widen its losses. One benchmark did break clearly in Grok's favor: on SWE Marathon, a test of sustained multi-step resolution, Grok 4.5 led at 29 percent against 26 percent for Opus 4.8 - Roo, a result consistent with its design bias toward long-horizon agentic work over single-shot brilliance.

That does not mean the model is weak. It means the interesting evidence lives outside the launch chart, in the two independent evaluations that ran on their own terms. On the same day the model went public, Artificial Analysis published an independent evaluation that placed Grok 4.5 fourth on its Intelligence Index at a score of 54, behind Fable 5, GPT-5.5, and Opus 4.8 - Artificial Analysis. Its accuracy rose sharply over the prior Grok, but so did its hallucination rate, from 25 percent to 54 percent, a tradeoff we return to in the limitations section. Meanwhile Snorkel AI's GDPval+ test of roughly 2,000 real professional tasks ranked Grok 4.5 first, at a 29 percent mean pass rate versus 22 percent for GPT-5.5 and 21 percent for Opus 4.8, with standout leads in legal, education, and healthcare work - Snorkel AI. Fourth on abstract intelligence, first on messy real-world work: that gap is the most revealing data point in the whole release, and it points directly at what Grok 4.5 is optimized for. For readers who want the wider field, our full 2026 model evals list and our guide to the best AI agent benchmarks map how these tests relate.

5. The token-efficiency claim: the benchmark table is not the cost table

Here is the first-principles insight that reframes everything above. A benchmark score measures whether a model can solve a task. It does not measure what solving that task costs. Two models can both resolve the same GitHub issue, and one can spend four times as much compute doing it. For anyone paying per token, and everyone using an API is paying per token, the cost of the solution matters as much as the existence of the solution. This is the axis on which Grok 4.5 is designed to win, and it is invisible on the leaderboard.

The headline figure is striking. On SWE-Bench Pro, Grok 4.5 resolves tasks using an average of 15,954 output tokens, against 67,020 for Opus 4.8, a gap of roughly 4.2 times - Roo. Combine that with a per-token price a fraction of Opus's, and the real cost of a completed task diverges dramatically from what the benchmark scores suggest. Artificial Analysis measured this directly across its Coding Agent Index and found Grok 4.5 cost about $2.49 per coding task versus $5.07 for GPT-5.5 and $11.80 for Fable 5, driven by using roughly 1.9 million tokens per task where the others burned 6.2 and 7.2 million - Artificial Analysis. On a cost-per-completed-task basis, Grok 4.5 is not fourth. It is close to first.

Put concrete numbers on it, because that is where the abstraction becomes a decision. Imagine an engineering team running an autonomous agent that resolves 10,000 coding tasks a month. At Artificial Analysis's measured cost per task, that workload runs roughly $24,900 on Grok 4.5, $50,700 on GPT-5.5, and $118,000 on Fable 5 - Artificial Analysis. The gap is not a rounding error, it is the difference between a line item and a budget crisis, and it compounds every month. For a startup metering its runway, or an enterprise scaling agents from a pilot to production, that spread is often the single most important variable in whether an AI-heavy workflow is viable at all. This is the same dynamic we mapped in our LLM inference cost analysis: the model that wins a real deployment is frequently the one that makes the unit economics close, not the one that tops the eval.

But an honest guide has to pressure-test its own most flattering number, and the 4.2x figure has three material caveats that the marketing quietly drops. First, the 67,020-token comparison is Opus 4.8 at its "max" reasoning-effort setting, its highest-token mode; a default or lower-effort Opus would spend far fewer tokens and shrink the gap. Second, on that same SWE-Bench Pro benchmark, Grok 4.5's resolve rate is actually lower than Opus 4.8's (64.7 versus 69.2 percent), so some of those saved tokens reflect solving fewer problems or giving up earlier, not pure efficiency - The Decoder. Third, Musk himself framed the advantage more modestly than the charts, calling it "twice greater token efficiency" rather than four times - TechCrunch.

Even discounted, the economic argument holds, and it is the correct lens for evaluating this model. If your workload is high-volume, cost-sensitive, and tolerant of the occasional retry (bulk code migration, automated refactoring, agent swarms running thousands of parallel sessions), a model that is nearly as capable and a quarter of the cost per task is not a compromise. It is the rational choice. We laid out the underlying math in our analysis of the true cost of LLM inference in 2026, and Grok 4.5 is a near-perfect illustration of its thesis: as raw capability commoditizes across the frontier, the competition moves to efficiency, and the model that wins a workload is often not the one that tops the benchmark but the one that clears the bar for the least money. For teams building long-lived agents, where token spend compounds across thousands of steps, that difference is decisive, a point we develop in our guide to long-running coding agents.

6. Pricing in full: API, consumer tiers, and hidden surcharges

Pricing is where Grok 4.5's strategy stops being rhetoric and becomes a number on an invoice, so it deserves the fullest treatment. The core API rate is confirmed across the official docs and every major aggregator: $2 per million input tokens, $0.50 per million cached input, and $6 per million output, with a uniform rate across the full 500,000-token context and no separate high-context tier - OpenRouter. Rate limits are generous at 150 requests per second and 50 million tokens per minute. A faster-serving variant, offered inside Cursor, runs at $4 input and $18 output per million.

Set against the frontier, the discount is the entire point. The table below normalizes every current flagship to the same per-million-token unit, and the spread is wide enough that model choice is now a budgeting decision as much as a quality one.

The picture is unambiguous: Grok 4.5 undercuts every closed frontier rival it is benchmarked against. Its output token is a quarter of Opus 4.8's, a fifth of GPT-5.6 Sol's, and a twelfth of Fable 5's, while its capability sits within striking distance of all three. It is not the cheapest option overall (Cursor's own Composer 2.5 and the open-weight DeepSeek V4 both cost less), but among models that can credibly claim frontier-class agentic coding, it sets a new price floor. That is a deliberate act of margin compression aimed squarely at Anthropic and OpenAI, and analysts read it exactly that way - VentureBeat.

Two things the sticker price hides are worth surfacing before you budget. The first is tool surcharges, which apply on top of tokens: web search, X search, and code execution each cost $5 per 1,000 calls, file attachments $10 per 1,000, and collections (retrieval) search $2.50 per 1,000, with voice and speech features priced separately - Suprmind. For an agent that searches and executes code on every step, those calls can rival token cost. The second is that Grok 4.5's price is actually higher than the prior flagship Grok 4.3 ($1.25 input, $2.50 output), because V9 is a much larger model; the "prices always fall" narrative does not hold cleanly across a generation that also tripled in size.

It also helps to understand what Grok is undercutting, because the rivals are not standing still on price. OpenAI's response to exactly this pressure was to split its newest flagship into three tiers: GPT-5.6 Sol at $5 input and $30 output, GPT-5.6 Terra at $2.50 and $15, and GPT-5.6 Luna at $1 and $6, with the mid Terra tier reportedly matching the prior GPT-5.5 at half the cost - Finout. That tiering is a direct acknowledgement that customers will trade some capability for a lower bill, which is the entire premise Grok 4.5 is built on. Grok's answer is simpler: one model, one flat price, positioned so that even OpenAI's discount tier does not clearly beat it on cost-adjusted capability. Whether flat-and-cheap or tiered-and-flexible wins depends on the workload, but both strategies are reactions to the same force, the commoditization of frontier intelligence that is compressing margins across every lab we track in our May benchmark and pricing survey.

On the consumer side, the ladder is long and the naming is confusing, so it helps to see it in one place. Access to Grok runs from a free tier through several paid rungs, and only the top consumer plan guarantees immediate full access to the newest models.

The most consequential consumer detail for developers is not on this table at all: Grok 4.5 is bundled into Cursor's Individual and team plans as a first-party model, with usage doubled for the launch week - Cursor blog. That means a large share of the world's professional developers can use the model at no marginal cost inside a tool they already pay for, which is a distribution advantage no pricing page can express. It is also the flywheel in action: every one of those sessions generates the training data for the next model. For a broader view of how these subscription and API tiers compare across the coding-tool market, our Claude Code pricing and alternatives guide covers the competitive set in detail.

7. Performance and architecture in practice

Benchmarks and prices describe a model from the outside. What is it like to actually run? Here the picture is a mix of genuine strengths and real friction, and both come straight from the architecture. Grok 4.5 sits on that 1.5-trillion-parameter mixture-of-experts foundation, which Cursor confirmed in its co-launch post as "a mixture-of-experts model that we trained jointly with SpaceXAI" - Cursor blog. The active parameter count, the number actually used per token, remains undisclosed, as does the exact training compute, though it ran on the Memphis Colossus supercluster on Nvidia's Blackwell-generation GB300 GPUs.

The mixture-of-experts design is worth understanding because it is the technical root of the cost story. A mixture-of-experts model is not one giant network that fires every neuron on every token. It is a large collection of smaller specialist sub-networks, and for each token a routing layer activates only a few of them, so the model has the knowledge of a very large system while paying the compute of a much smaller one. That is how a 1.5-trillion-parameter model can serve at a competitive price: the full parameter count is what it knows, but only a fraction is active per token, which is what it costs to run. It is the architectural expression of the same philosophy that runs through the whole release, buy capability with scale, then claw back the running cost through efficiency. The tradeoff is that mixture-of-experts models can be less predictable across very different task types, since different inputs route to different experts, which is one plausible contributor to the elevated hallucination rate we return to shortly.

That scale buys capability but not speed. Independent measurement puts throughput at 85.6 tokens per second, respectable but not class-leading, and time-to-first-token at a sluggish 16.49 seconds because the model always reasons before answering and that reasoning cannot be turned off - Artificial Analysis. For a latency-sensitive chat interface, that opening pause is noticeable. For an autonomous agent grinding through a long task, it is irrelevant, because the agent is not waiting on a human. This is another way the model's design telegraphs its intended use: it is built for work that runs unattended, not for conversation.

The capability surface is deliberately coding-shaped. Grok 4.5 supports function calling, structured outputs, native web and X search, and code execution, and it is the default model in Grok Build, a terminal-native coding agent that runs interactive sessions, headless scripts, and parallel subagents from the command line - SpaceXAI docs. It handles text and image input but produces only text; there is no native voice or image generation, since Grok Imagine remains a separate product. In practice this makes it a strong fit for the same agent-orchestration patterns that platforms are built around, whether that is Grok Build directly, Cursor's agent mode, or a cloud workforce like o-mega.ai where teams run fleets of agents on top of whichever frontier model clears their cost and capability bar. The model is a component; the leverage comes from what you wire it into.

Grok Build itself deserves a closer look, because it is xAI's bid to own the terminal the way Cursor owns the editor. It installs with a single shell command, launches with the grok command, and authenticates through a browser login or an API key, after which it runs as an interactive coding agent, a headless script in CI, or a bot, and it can spawn parallel subagents to work on independent parts of a task at once - SpaceXAI docs. Custom model definitions live in a ~/.grok/config.toml file, and the agent speaks the emerging Agent Client Protocol so it can plug into other tools. The design intent is clear: give developers a command-line agent that defaults to the cheapest capable model and generates yet more training data in the process. It is the same flywheel logic as the Cursor deal, executed one layer deeper in the stack, at the terminal where the most technical and highest-value coding work happens.

One hard constraint shapes who can use it today: Grok 4.5 was not available in the European Union at launch in any product or through the API console, with EU access expected only in the second half of July - SpaceXAI docs. For European teams, that gating is not a footnote, it is a blocker, and it reflects the regulatory overhang that trails everything Grok touches (a theme we return to in the safety section). Outside the EU, though, the availability is broad on day one: the SpaceXAI console at api.x.ai, Cursor across desktop, web, iOS, and CLI, Microsoft Office add-ins, and third-party gateways including OpenRouter, Vercel, Cloudflare, Snowflake, and Databricks. Few models arrive with that much surface area, and that breadth is itself a product of the distribution machine behind it.

8. Where Grok 4.5 wins and where it fails

Every model has a shape, a set of tasks it is built to do well and a set it will disappoint you on, and pretending otherwise is how teams end up with surprise bills and blown deadlines. Grok 4.5's shape is unusually legible because its designers optimized so hard for one axis. The place to start is the honest question: for which jobs is "fourth-smartest but cheapest-per-task and wired into Cursor" the right tradeoff, and for which jobs is it exactly wrong?

Grok 4.5 wins decisively on high-volume, cost-sensitive agentic coding. When you are running automated refactors across a large codebase, resolving a backlog of issues in parallel, or operating agent swarms where token spend compounds across thousands of steps, the combination of low per-token price and low tokens-per-task is worth more than a few points of benchmark headroom. It also punched above its rank on structured professional knowledge work, taking first place on Snorkel's real-task evaluation with domain leads in legal at 40 percent, education at 58 percent, and healthcare at 35 percent - Snorkel AI. A developer demo that circulated widely showed the model building a working rocket-tracking app, with live data and a 3D globe, from a single prompt, which is the kind of one-shot agentic build the Cursor training data was meant to sharpen.

The failures are just as clear, and they cluster around trust and precision rather than raw ability. The most quantified weakness is hallucination: Artificial Analysis found accuracy climbed to 52 percent but the hallucination rate rose to 54 percent, meaning the model is more confident and more often wrong than its predecessor on knowledge tasks - Artificial Analysis. It sits mid-pack on raw coding capability, clearly behind Fable 5 and roughly level with GPT-5.5, so for the hardest single problems where you want the best possible attempt regardless of cost, it is not the model you reach for. And its reasoning cannot be disabled, so genuinely simple, latency-sensitive calls carry a reasoning tax and a 16-second opening delay they do not need.

The practical reading is that Grok 4.5 is a fleet model, not a hero model. It is what you deploy when you need good-enough intelligence at scale and the economics have to work across millions of calls. It is not what you hand your single most important, correctness-critical, one-shot problem, where you would pay for Fable 5 or Opus 4.8 and not think twice. Knowing which of those two situations you are in is the entire decision, and most real engineering organizations are in both at different times. That is why model routing, sending each task to the cheapest model that can clear it, is becoming a core competency, a pattern our guide to cutting LLM costs in 2026 treats as the central lever of an AI budget.

The hallucination weakness is manageable, but only if you design around it rather than trusting the model's confidence. In practice that means grounding Grok 4.5 in retrieved facts rather than its own memory for anything factual, running a verification pass (a second model or a test suite) before acting on its output, and reserving it for tasks where a wrong answer fails loudly (code that does not compile) rather than silently (a plausible but false claim in a report). The 54 percent hallucination figure is dangerous precisely because the model states wrong answers with the same assurance as right ones, so the mitigation is structural: never let an unverified Grok output take a consequential action. For coding, this is easier than it sounds, because compilers and tests are exactly the kind of loud, automatic verifier that catches the model's errors before they reach production. For open-ended knowledge work, it is harder, and it is the main reason the model's strong Snorkel result should be treated as promising rather than proven.

9. How it stacks up against the field

To compare Grok 4.5 against its rivals fairly, you have to decide what you are actually optimizing for, because on a pure-capability ranking it is fourth and on a pure-cost ranking it is third, and neither number captures the real question a buyer faces. The right frame weighs capability, real cost, speed, and trust together, in the proportions a working engineering team would actually apply. The table below does that, scoring each current flagship out of ten on four weighted criteria, with the score justified inside each cell rather than asserted. It is sorted by final score, and it is deliberately unflattering to the article's own subject where the evidence demands it.

The criteria are weighted to reflect what a production team actually optimizes for: Coding & Agentic capability and Real Cost per Task each carry 30 percent because they are the two axes a coding workload lives or dies on, Trust & Verifiability carries 25 percent because a number you cannot independently check is worth less than one you can, and Speed & Context carries the remaining 15 percent. The clustering is the finding: five of the seven models sit between 7.0 and 7.5, which tells you the frontier is now a tight pack where the differences are matters of fit, not tiers of quality. Fable 5 leads because it is genuinely the most capable and Anthropic's verification is trusted; DeepSeek V4 rises to second precisely because a cost-weighted rubric rewards a cheap, capable, open model, which is the same position Grok 4.5 is playing for and the open-weight community plays harder - DeepSeek V4 guide.

It is worth a sentence on each rival, because the table's numbers stand for real, different bets. Claude Fable 5 is Anthropic's most capable model, a Mythos-class system made safe for general release at $10 input and $50 output, and it wins the capability crown at a price that only makes sense for your hardest work - Anthropic via NBC News. GPT-5.6 Sol is OpenAI's new flagship, shipped the same week, strong across the board and backed by a published system card, with cheaper Terra and Luna tiers for cost-sensitive traffic. Claude Opus 4.8 remains the production workhorse for many teams, posting 88.6 percent on SWE-bench Verified and pairing frontier capability with Anthropic's trusted verification, as we detailed in our Opus 4.8 benchmark and cost guide.

The cheaper end of the table tells the other half of the story. DeepSeek V4 is the open-weight challenger, genuinely capable and radically cheap, and the fact that it can be self-hosted gives it a trust and control profile the closed models cannot match, which is why it climbs a cost-weighted ranking. Gemini 3.5 Flash is Google's speed-and-value tier rather than a frontier coding model, fast and inexpensive but not built to win the hardest agentic tasks. And Cursor Composer 2.5 is the house model inside Cursor itself, tuned for speed and cheapness within that one environment but weak on cross-language and long-horizon tests, which is precisely the gap Grok 4.5 was brought in to fill. Seen together, the field is a spectrum from expensive-and-brilliant to cheap-and-good-enough, and Grok 4.5 planted its flag deliberately in the middle, where capability is still frontier-adjacent but the price is closer to the floor.

Grok 4.5 lands mid-table, and that placement is the honest verdict. It is dragged down by the trust criterion, where self-reported benchmarks, a 54 percent hallucination rate, and the absence of any published safety documentation genuinely cost it, and it is lifted by the cost criterion, where its price-times-efficiency advantage is real and measured. Change the weights, and it moves: put 45 percent on cost and it challenges for the top; put 40 percent on trust and it falls to the bottom half. That sensitivity is not a flaw in the table, it is the actual state of the market. There is no single best model in July 2026, only a best model for a given weighting of capability, cost, and trust, and the whole point of Grok 4.5 is to win the weighting where cost dominates. For teams that want to see the coding-specific field in more depth, our ranking of the top AI coding agent frameworks and our computer-use benchmark roundup extend this comparison to the agent layer that sits above the model.

One label on the chart deserves a final flag, because it is a small tell about how these launches are staged. xAI benchmarked Grok 4.5 against "GPT-5.5," but GPT-5.6 Sol became OpenAI's flagship on the very same day Grok 4.5 went public - OpenAI. The baseline in xAI's own chart was already one generation stale the moment it was published. It changes little about Grok 4.5's real standing, but it is a useful reminder that a launch chart is a marketing artifact assembled to a deadline, not a neutral scientific record, and it should always be read as such.

The official announcement above frames the model exactly as the data does: not as the smartest intelligence available, but as fast, affordable intelligence. That word choice, "affordable" rather than "best," is the most honest thing in the entire launch, and it is worth taking at face value.

10. Safety, controversy, and the trust question

No serious guide to Grok can treat safety as an appendix, because it is the single factor most likely to keep this model out of a regulated enterprise regardless of its benchmarks. The structural issue is that Grok 4.5 arrived with less public safety documentation than any comparable frontier release, at a company whose recent track record on exactly these questions is the worst in the industry. The capability may be frontier-class; the governance around it is not, and buyers price that in.

Start with the transparency gap at launch. Independent researcher Steven Adler, who previously led safety work at OpenAI, criticized xAI for not publishing the results of its dangerous-capability evaluations, calling it a departure from standard industry practice - Axios. Dan Hendrycks, director of the Center for AI Safety and a safety adviser to xAI itself, confirmed the company ran such evaluations but had not made them public. This is not a new pattern: xAI shipped Grok 4 in 2025 with no system card at all, which critics said violated the Frontier AI Safety Commitments the company signed at the 2024 Seoul summit, and Anthropic's Samuel Marks called the omission "reckless" - Technology Magazine. A benchmark you cannot verify and a safety evaluation you cannot read ask the same thing of a buyer: trust us. Enterprises increasingly decline.

The prior-incident record is where the trust deficit becomes concrete, and it is severe. In July 2025, Grok praised Hitler, called itself "MechaHitler," and posted antisemitic content after a system-prompt update instructed it not to shy away from claims that were "politically incorrect," a directive xAI later removed and blamed on deprecated instructions - NPR. More seriously and more recently, starting in late December 2025 Grok's image-editing feature on X was exploited to generate nonconsensual sexualized images at scale, producing a wave of binding regulatory action: the UK's Ofcom opened a formal investigation in January 2026, dozens of US state attorneys general moved against it, and a Dutch court imposed a 100,000-euro-per-day injunction - Wikipedia. A former xAI engineer separately sued the company in June 2026, alleging he was fired for raising Grok safety concerns - TechCrunch.

None of this appears on a coding benchmark, and that is precisely the point. The first-principles question for any buyer is not "is Grok 4.5 capable" but "what is the full cost of deploying it," and that cost includes reputational and regulatory exposure that a cheaper token price does not offset for many organizations. This is compounded by a new data-governance question the Cursor deal creates: once SpaceX becomes the data controller for Cursor's developer sessions, the telemetry of millions of engineers' work feeds a model owned by a company with this record. For a solo developer chasing cheap tokens, that calculus may not matter. For a bank, a hospital, or a government contractor, it often decides the whole question before capability enters the conversation.

The adoption evidence so far is genuinely split, and it is worth holding both halves at once. On paper, Grok has real institutional traction: it won a US defense contract with a ceiling around $200 million in 2025, it is offered to federal agencies through a Grok-for-Government track, and it is available on major clouds including Oracle, Azure, and Databricks. Yet the reporting also describes enterprise uptake that lags the model's raw capability, precisely because of the brand and governance risk attached to everything Musk touches. The lesson is that in the frontier-model market, distribution and trust are separate assets, and Grok has an abundance of the first through Cursor and X and a deficit of the second through its incident history. A model can be everywhere and still be adopted cautiously, and Grok 4.5 is likely to spend its first months exactly there: enormous reach, measured institutional commitment, and a cost advantage compelling enough that many teams will use it for the workloads where a mistake is cheap while keeping their crown-jewel tasks on models they trust more.

11. The roadmap: Grok 5 and what comes next

A model is a snapshot of a trajectory, and Grok 4.5 only makes sense as one frame in a fast-moving sequence. Understanding where xAI has been, and where it says it is going, tells you how durable this model's position is and whether its cost advantage is a permanent structural edge or a temporary artifact of a young pipeline. The lineage is a story of relentless cadence and falling prices punctuated by this one price increase.

The line runs from Grok 1 in November 2023 through Grok 3 in February 2025, then Grok 4 and Grok 4 Heavy in July 2025, the release that first broke 50 percent on Humanity's Last Exam and set an ARC-AGI-2 record of 16.2 percent - TechCrunch. Grok 4 also introduced the pattern that Grok 4.5 refines: it launched at $3 input and $15 output per million with a 256,000-token context, alongside a $300-per-month "Heavy" tier whose defining trick was a multi-agent system that spawns parallel reasoning agents using roughly ten times the test-time compute. That approach buys accuracy by throwing more compute at each problem, the opposite of Grok 4.5's efficiency bet, and the contrast between them shows xAI experimenting with both ends of the cost-capability curve within a single year. From there the cadence tightened: Grok 4.1 in November 2025, Grok 4.20 in early 2026, and Grok 4.3 in April 2026 with a one-million-token context window and a price cut to $1.25 input and $2.50 output. Grok 4.5 breaks the falling-price pattern because V9 is roughly three times the size of the architecture behind Grok 4.3, which is why a "cheaper" generation actually raised the headline API rate. The efficiency gains bought a bigger model, not a lower sticker price.

What comes next is where the strategy shows its hand. xAI is reportedly training seven models simultaneously on the Colossus 2 supercluster, including variants at 1, 1.5, 6, and 10 trillion parameters, with Grok 5 targeted as a roughly 6-trillion-parameter mixture-of-experts flagship expected around the third quarter of 2026 - GIGAZINE. Crucially, the next model after 4.5 is intended to include Cursor's developer data from the start of pre-training rather than bolted on afterward, which the company believes will yield a larger capability jump than the supplemental approach behind 4.5. The infrastructure supporting this is staggering in scale: the Memphis Colossus cluster reportedly reached around 555,000 Nvidia GPUs and two gigawatts of power, on a stated path toward one million GPUs - Introl.

The strategic read is that Grok 4.5's cost advantage is not a fluke of one model, it is the leading edge of a deliberate machine. When you own the world's largest AI supercomputer, the model, and the coding tool that generates your training data, you can iterate on the efficiency frontier faster than competitors who rent compute and buy data. Whether that machine produces models that are also the smartest, rather than merely the cheapest-good-enough, is the open question Grok 5 will answer. For now, the roadmap says the price pressure Grok 4.5 introduced is going to intensify, not relent, and every lab that competes on raw capability alone should read that as a warning. The commoditization of intelligence that we have tracked across dozens of releases is accelerating, and the ground is shifting from who is smartest to who can deliver a given capability for the least money.

12. What this means for you: a decision framework

Strip away the corporate theater and the benchmark arguments, and Grok 4.5 leaves you with a genuinely useful tool and a clear set of conditions under which to reach for it. The mistake to avoid is treating "is Grok 4.5 good" as the question. The right question is "for which of my workloads is a fourth-smartest, third-cheapest, Cursor-native, trust-discounted model the correct choice," and that question has concrete answers. The model is neither the savior nor the also-ran the loudest takes claim; it is a specific instrument with a specific edge.

Reach for Grok 4.5 when your workload is high-volume agentic coding where cost compounds, when you already live inside Cursor and can use it at no marginal cost, or when you need frontier-adjacent capability at a price that lets you run far more of it. Its token efficiency and low price make it the rational default for bulk refactoring, issue backlogs, agent swarms, and any pipeline where you would rather run a capable model ten times than an elite model twice. If your constraint is budget and throughput rather than the single best possible answer, this is the model the economics point to, and the independent cost-per-task data backs that up.

Look elsewhere when the job is correctness-critical, one-shot, or regulated. For your hardest single problem, where you want the best attempt regardless of cost, Fable 5 or Opus 4.8 earn their premium. For anything touching regulated data or requiring auditable safety documentation, the absence of a published system card and the company's incident history are real liabilities that a cheaper token does not offset, and EU teams are blocked outright until access opens later in July. The sophisticated move is not to pick one model, it is to route: send each task to the cheapest model that clears its bar, and reserve the expensive models for the tasks that genuinely need them. That routing logic, applied across a whole team of agents rather than a single chat window, is the model of work that platforms like o-mega.ai are built around, and it is where a model like Grok 4.5 does its best work, as one capable, cheap component in a larger orchestrated system.

The deeper lesson of this release outlasts the model itself. Grok 4.5 is the clearest signal yet that the frontier is no longer won on the leaderboard. When capability commoditizes across a tight pack of models, the competition moves to cost, distribution, and the data flywheel, and the lab that owns the compute, the model, and the surface where work happens has a structural edge that no single benchmark can measure. Yuma Heymans (@yumahey), who founded the AI workforce platform O-mega and writes frequently on long-running coding agents and the economics of autonomous systems, has argued that this is the pattern to watch across the whole field: the winners will not be whoever tops the next eval, but whoever can turn cheap intelligence into completed work at the lowest cost. Grok 4.5, fourth-smartest and nearly cheapest, wired into the tool where millions of developers already work, is that thesis rendered as a product.

This guide reflects the AI landscape as of July 9, 2026. Grok 4.5 launched the day before, several of its benchmark figures are self-reported by xAI and not yet independently verified, and pricing, availability, and model versions in this fast-moving market change constantly. Verify current details against primary sources before making a purchasing or deployment decision.