Is Claude Fable 5 available (Access Guide July 2026)

The definitive guide to when, where, and for what Claude Fable 5 is actually available, updated for July 2026.

Claude Fable 5, Anthropic's most capable widely released model, was live for exactly three days before the US government pulled it off the internet. It launched on June 9, 2026, went dark worldwide on June 12 under an emergency export-control order, and came back globally on July 1 - Anthropic. If you have tried to answer the simple question "is Claude Fable available right now" over the past month, you have been chasing a moving target, because the honest answer changed four times in three weeks.

So here is the short version: yes, Fable 5 is available again as of July 1, 2026, but "available" is doing a lot of work in that sentence. Availability is not a single yes or no. It is a matrix. Whether you can use Fable 5 depends on where you are, what plan you pay for, how much you have already used this week, and what you are asking it to do. The same account can have Fable 5 for one prompt and be silently handed a different model on the next, and that is by design, not by accident.

This guide breaks the question apart into the dimensions that actually determine access: geography, platform and surface, subscription plan and quota, and task domain. It walks through the 19-day shutdown that made this the most contested model launch of the year, explains the safety architecture that decides which of your requests Fable 5 will and will not answer, corrects the pricing myth that half the internet is repeating, and closes with the specific alternatives to reach for when Fable 5 is not the model you can (or should) use. The goal is that by the end you can look at your own situation and know, without guessing, whether Fable 5 is available to you and what to do when it is not.

Contents

1. The short answer: available, with strings attached
2. What Claude Fable 5 is, and why it is not Mythos
3. The 19 days Fable vanished: a timeline of the shutdown
4. Availability by geography: global again, but read the fine print
5. Availability by platform and surface: API, clouds, apps, and Code
6. Availability by plan and quota: who pays, who waits, who cannot
7. Availability by task: the safety classifiers and the Opus 4.8 fallback
8. Mythos 5 and Project Glasswing: the version almost nobody can access
9. What Fable 5 actually costs, and the pricing myth to ignore
10. Why availability matters: capability, benchmarks, and lock-in
11. When Fable 5 is not your answer: the alternatives, scored
12. How to build so a single model outage cannot stop you

1. The short answer: available, with strings attached

The reason "is Claude Fable available" resists a clean answer is that availability is gated at four independent layers, and you have to clear all four to actually get a response from the model. Clearing three of them and failing the fourth still means no Fable 5. Most of the confusion online comes from people answering one layer ("yes, it is back on Claude.ai") as if it settled all four. It does not. A researcher in Berlin, a startup on the free tier, a user who has burned through their weekly allowance, and a security engineer asking about exploit code can all be told "not available" for four completely different reasons on the same afternoon.

The first principle worth holding onto is that Anthropic treats Fable 5 as a controlled good, not a consumer product. Everything downstream, the export order, the nationality checks, the mandatory data retention, the automatic model switching, follows from the company's own claim that this model class is powerful enough to matter to national security - TechCrunch. Once you accept that framing, the strange availability rules stop looking arbitrary and start looking like the deliberate friction of a company shipping something it considers genuinely dangerous. That is the lens for the entire guide.

Here is the availability matrix at a glance. Each row is an independent gate, and the "condition" column is where the real answer lives.

The table looks busy on purpose, because the busyness is the point. To make it concrete, the diagram below is the same logic as a decision path. Walk it top to bottom and you will land on exactly one of five outcomes, which is a far more useful mental model than "is it up."

What the diagram makes obvious is that the most common "failure" is not a hard block at all. It is the bottom branch, where a request quietly gets answered by Claude Opus 4.8 instead of Fable 5, and the user is told after the fact. That is not the model being down. That is the model working as specified. Understanding that distinction is the single most important thing in this guide, and section 7 unpacks exactly how it works. For now, hold the frame that Fable 5 availability is conditional, layered, and mostly working, with a handful of well-defined exceptions that you can plan around.

2. What Claude Fable 5 is, and why it is not Mythos

Before you can reason about availability, you need to know what is actually being made available, because Anthropic shipped two models on June 9, not one, and they are the same brain wearing different amounts of armor. Claude Fable 5 (model id claude-fable-5) is the public model. Claude Mythos 5 (claude-mythos-5) is the same underlying model with the safety classifiers removed - Claude Platform Docs. Everything that makes Fable 5 feel restricted is the armor. Mythos 5 is what the model can do with the armor off, and Anthropic has decided the public does not get that version.

The naming is not marketing whimsy, it is a tell about the strategy. Anthropic chose "Fable" deliberately: it comes from the Latin fabula, "that which is told," an echo of the Greek mythos - Anthropic. Fable is, quite literally, the version of the myth that gets told to everyone. Mythos is the original. The company has been explicit that no one has yet built safeguards strong enough to release a Mythos-class model to the public without the classifiers, which is the whole reason Fable exists as a separate product line - Anthropic. If you internalize nothing else about the product architecture, internalize that Fable is Mythos plus a set of filters, and those filters are the availability story.

On raw specifications the two are identical, and the numbers are genuinely large. Both carry a 1 million token context window by default and up to 128,000 output tokens per request, with a knowledge cutoff of January 2026 - InfoQ. Both are built for what Anthropic calls "long-horizon agentic work," meaning tasks that run for hours across many tool calls rather than single question-and-answer turns. A few behaviors are specific to this model class and worth knowing before you commit to it:

• Adaptive thinking is always on and cannot be disabled, so you control depth with an effort parameter rather than a toggle.
• Raw chain-of-thought is never returned, only a summarized version or nothing, which matters for anyone who relied on reading Claude's reasoning.
• File-based memory helps Fable disproportionately, reportedly around three times more than it helped Opus, because long-horizon work leans on persistent state.
• Vision, code execution, and programmatic tool calling ship at launch, so it is a full agentic model, not a text-only endpoint.

Those design choices explain why Fable 5 is not simply "a bigger Opus." It is tuned for autonomy: it thinks by default, it remembers across a task, and it expects tools. That is also why the availability rules are stricter than for older models, because an always-thinking, tool-using, long-running agent with frontier cybersecurity and biology skills is exactly the artifact governments worry about. The capability and the caution are two sides of the same coin, and the rest of this guide is about the caution.

It is worth dwelling on what long-horizon agentic work actually demands, because it explains the rest of the product. A model that runs for hours cannot re-read its entire history on every step, so persistent memory and context management stop being conveniences and become the difference between a coherent agent and one that forgets its own plan halfway through. That is why the always-on thinking and the file-based memory matter far more here than on a chat model: they are the machinery that lets Fable hold a goal across thousands of tool calls. It is also why the summarized-only thinking output is a deliberate trade. Anthropic will let you see that the model reasoned, and a summary of how, but not the raw chain, both to protect the model from distillation and to keep the reasoning itself inside the safety perimeter.

For a deeper capability comparison against its sibling and predecessors, our Fable 5 and Mythos 5 benchmark breakdown goes number by number, and the broader Anthropic ecosystem guide situates both models in the full lineup.

3. The 19 days Fable vanished: a timeline of the shutdown

You cannot understand today's availability rules without understanding the crisis that produced them, because the current 50-percent caps, nationality clauses, and 99-percent classifiers are all direct scar tissue from June. The story is short, strange, and unprecedented for a commercial AI model. On June 9 Anthropic shipped Fable 5 and Mythos 5. On June 12 the US Commerce Department ordered both offline. On June 30 it lifted the order. On July 1 Fable came back. In between, the most capable model available to the public simply did not exist for anyone, anywhere, for the better part of three weeks - CNBC.

The trigger was a jailbreak. Researchers at Amazon found a prompt sequence that bypassed Fable 5's safeguards and got the model to identify software vulnerabilities, and in at least one case to write code demonstrating how a flaw could be exploited - The Hacker News. Amazon's leadership escalated the finding to the federal government, and the response was extraordinary: rather than a fine or a warning, the Commerce Department invoked export-control authority. On the evening of June 12, at 5:21pm Eastern, Anthropic received a directive giving it roughly ninety minutes to comply - Fortune.

The directive's scope is the part that broke everyone's mental model of how AI access works. It did not ban a country. It banned a category of person. The order suspended all access to Fable 5 and Mythos 5 by "any foreign national, whether inside or outside the United States, including foreign national Anthropic employees" - Anthropic. This is the "deemed export" doctrine from traditional technology controls applied to a chatbot: a US citizen in Berlin was still permitted, while a foreign national in San Francisco was not. Because Anthropic had never built real-time nationality verification into its platform, it could not enforce a person-by-person rule, so it did the only compliant thing and disabled both models for every user on Earth - Forbes. Opus, Sonnet, and Haiku stayed up the whole time, which tells you the government saw Fable and Mythos specifically as the sensitive artifacts.

The deemed-export doctrine is worth pausing on, because it is the mechanism that turned a software update into a national-security incident. In traditional export law, showing controlled technology to a foreign national is legally equivalent to shipping it abroad, even if the person never leaves the country. Applied to a model, that logic means a query typed in San Francisco by someone on a work visa counts as an export to their home nation. Anthropic had built its platform on the ordinary assumption that a user is a user, so it had no switch to enforce a rule written in the language of 1990s hardware controls. The ninety-minute compliance window left no time to build one, and the only lever that guaranteed compliance was the off switch. That is how a jailbreak in a research lab became a global outage.

The blast radius reached inside Anthropic itself, which is the detail that made the episode feel genuinely new. Because the order named foreign-national employees explicitly, the company's own non-citizen staff were, on paper, barred from the models they had helped build. Downstream partners moved even faster than the formal order: Microsoft pulled Fable 5 from its internal Copilot on June 10, two days before the directive, once the risk became visible - InfoQ. The lesson enterprises took away was not about one vendor. It was that a model you depend on can be removed by a government that is not your vendor, on a timeline you do not set, for a reason you cannot appeal.

The visual below lays the sequence out. Note that Mythos 5 came back for a narrow set of US organizations on June 26, four days before the general public got Fable back, which is itself a clue about who Anthropic prioritizes.

The resolution came with conditions attached, and those conditions are now permanent features of the product. To get the order lifted, Anthropic trained a new cybersecurity classifier that blocks the specific Amazon technique in over 99 percent of cases, and it agreed to deeper cooperation with the government on pre-release testing, information sharing, and research collaboration - Anthropic. It also proposed, with Amazon, Microsoft, and Google, an industry-wide framework for scoring the severity of any future jailbreak along four axes: capability gain, breadth, ease of weaponization, and discoverability. In other words, the price of Fable 5 being available at all was a new layer of filtering and a new relationship with Washington.

It is worth noting that the severity was contested even as the order stood. Independent cybersecurity researchers argued the demonstrated capability was not unique to Fable 5 and could be reproduced with other publicly deployed models, a point Anthropic itself made when disputing the framing - CyberScoop. That debate matters for availability because it foreshadows the future: if regulators can pull a frontier model off the internet in ninety minutes over a contested jailbreak, then "available" is now a politically contingent state for every top-tier model, not just this one. We put that episode in the wider context of state control over models in our guide to sovereign AI and nationally owned models, which is suddenly a lot more than a theoretical topic.

4. Availability by geography: global again, but read the fine print

Geography is where the most confident wrong answers live, so it is worth being precise. As of the July 1 relaunch, Fable 5 is available to users globally, in Anthropic's own words, with no country-specific carve-outs for the model itself - Anthropic. Some early coverage claimed the relaunch was "US only," and that claim is simply wrong: it is contradicted by the primary source and by every outlet that read the actual statement rather than a summary of it - Al Jazeera. If you have seen "US only" repeated somewhere, that is a stale artifact of the June shutdown, not the current reality. Fable 5 today follows the same geographic rules as the rest of Claude.

Those rules are Anthropic's standard Supported Regions Policy, which the company describes as covering roughly 195 countries, regions, and territories - Anthropic. "Global" here means "everywhere Anthropic already operates," not "literally everywhere." The exclusions are the usual sanctioned and restricted jurisdictions, and they are consistent with US export law. Practically, the list of places where Claude, and therefore Fable 5, is not offered includes the following, which will surprise no one who follows sanctions:

• China, Russia, and Belarus, which are absent from the supported list entirely.
• Iran, North Korea, Cuba, and Syria, the long-standing comprehensively sanctioned set.
• Occupied regions of Ukraine, with Anthropic supporting Ukraine except Crimea, Donetsk, and Luhansk on its developer docs - Claude Platform Docs.
• Several conflict-affected African states, which appear on neither the platform nor the marketing region lists.

Beyond the country list, there is an ownership clause that enterprises should read carefully, because it can catch a company headquartered in a perfectly supported country. Anthropic reserves the right to refuse service to entities whose majority direct or indirect ownership is attributable to a non-supported nation, regardless of where the entity itself sits - Anthropic. A subsidiary in London owned mostly by a parent in a sanctioned jurisdiction is not automatically clear. For most readers this is irrelevant, but for multinationals with complex ownership it is the kind of detail that turns "available" into "available pending review."

A worked example makes the nationality-versus-residency distinction concrete, because it is genuinely counterintuitive. During the June window, a US citizen working remotely from Lisbon kept access, while a French national sitting in a San Francisco office lost it, even though the second person was physically closer to Anthropic's headquarters and inside US jurisdiction. Availability was pinned to the passport, not the postal code. That inversion is why Anthropic could not solve the problem with a geo-fence and had to pull the plug entirely. It also means that if this kind of order ever returns, a VPN would not restore access, because the control was never about where the traffic originates. It was about who is asking. The practical guidance for a compliance team is therefore to treat frontier-model access as a controlled-technology question, not just a data-privacy one. If your workforce includes foreign nationals and your product depends on a top-tier model, you now carry an export-control exposure that did not exist for older models, and it belongs in your risk register alongside data residency and retention.

The last geographic wrinkle is a pricing one, and it is easy to miss. If your compliance posture requires that inference physically stay inside the United States, Anthropic offers that, but it is not free. Setting inference_geo: "us" (or the equivalent US data-zone deployment on the marketplaces) keeps processing in-country at a 1.1x price multiplier across every token category - Claude Platform Docs. So geography touches availability twice: once as a hard gate (are you in a supported region) and once as a soft cost lever (do you need US-only residency, and will you pay ten percent more for it). Both are worth confirming before you build a workload on Fable 5, because unlike the June nationality rules, these are stable and unlikely to change.

5. Availability by platform and surface: API, clouds, apps, and Code

The surface you access Fable 5 through changes not just how you reach it but, in some cases, whether you can reach it at all and which features come with it. At the June 9 launch, Fable 5 was generally available across five developer platforms: the first-party Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud through Vertex AI, and Microsoft Foundry - Claude Platform Docs. That breadth is deliberate, because Anthropic wants Fable to be the default frontier model wherever developers already build. The consumer surfaces followed the same pattern: Claude.ai on web, desktop, and mobile, plus Claude Code and Claude Cowork.

The relaunch, however, did not restore every surface simultaneously, and that staggering still matters as of early July. When the controls lifted, the first-party surfaces came back first: Claude.ai, the Claude Platform (the API), Claude Code, and Claude Cowork were all live on July 1 - Anthropic. The three partner clouds were described as being re-enabled "as quickly as possible," without a firm date. So if you are an enterprise whose entire stack runs through Bedrock or Vertex, your "is it available" answer in the first days of July was genuinely different from a developer hitting the Anthropic API directly. Always check the specific cloud console rather than assuming parity with the API.

Even within the clouds, the regional footprint is narrower than most models, which trips up teams used to wide availability. On Amazon Bedrock, the in-region endpoints at launch were only US East (N. Virginia) and Europe (Stockholm), though a global inference profile spans roughly 35 regions by routing dynamically - Amazon Bedrock. Microsoft Foundry hosts Fable 5 only on the Anthropic-hosted global-standard deployment, not the Azure US data zone, which means a handful of Azure-native features do not apply to it. These are the sort of constraints that do not show up in a headline but absolutely show up in an architecture review, so treat "available on Bedrock" as the start of a question, not the end of one.

The regional narrowness has a practical consequence for latency and compliance that is easy to overlook. If your users sit in Asia-Pacific and Fable's in-region Bedrock endpoints are only in Virginia and Stockholm, you are either routing across an ocean or leaning on the global inference profile, which trades guaranteed data-region routing for availability. For a compliance regime that requires data to stay in a specific jurisdiction, that trade may not be acceptable, and you would fall back to a model with a broader regional footprint or to the first-party API with its US-residency option. None of this makes Fable unavailable, but it can make Fable the wrong availability shape for a given deployment, which is a subtler and more common problem than an outright block.

Two surface-specific details are worth calling out because they affect real workflows. First, in Claude Code, Fable 5 is available but is not the default model, and older clients cannot see it at all: you need version 2.1.170 or later, then select it with the model command, the --model claude-fable-5 flag, or the ANTHROPIC_MODEL environment variable - Claude Platform Docs. Second, Fable 5 also reached GitHub Copilot, but only for the Pro+, Max, Business, and Enterprise tiers, and it is off by default for organizations, requiring an administrator to enable it, partly because Fable's mandatory 30-day retention conflicts with Copilot's usual zero-retention posture - GitHub. The pattern across every surface is the same: Fable is broadly available but never quite frictionless, and the friction is always a safety or governance decision. Our Claude Code pricing and plans guide covers the Code-specific economics in more depth for anyone building there.

6. Availability by plan and quota: who pays, who waits, who cannot

For most people, "is Claude Fable available" really means "is it available on my subscription," and here the answer is the most conditional of all, because it is gated by both plan tier and a weekly consumption cap that Anthropic has never fully quantified. Start with the hard exclusions, because they are clean. Fable 5 is not on the Claude.ai Free plan, at all, ever. There is no complimentary path to the model on consumer surfaces - 9to5Mac. If you are on the free tier, Fable 5 is simply not one of your options, and no amount of waiting changes that. Access begins at the paid tiers.

Among paid users, availability after the relaunch is governed by a temporary and slightly awkward rule. For Pro, Max, Team, and select Enterprise plans, Fable 5 is included for up to 50 percent of your weekly usage limit through July 7, after which it moves to usage credits - Anthropic. In plain terms, you get to spend up to half of your normal weekly Claude budget on Fable during the grace window, and once you cross that line, or once July 7 passes, continued Fable use is billed as metered credits rather than folded into your flat subscription. The relevant consumer prices that gate this access are worth having in one place:

There are two traps in that table. The first is that "50 percent of your weekly usage limit" has never been published as a token or dollar figure by Anthropic, for any plan - Digital Applied. The underlying weekly ceilings are dynamic, so anyone quoting you an exact "X messages of Fable per week" number is estimating, not citing. The second trap is the burn rate. Anthropic's own in-app text warns that Fable 5 draws down your usage roughly twice as fast as Opus 4.8, partly because it thinks by default and generates longer outputs - Developers Digest. So that 50 percent allowance is consumed at double speed, and heavy agentic users on even the $200 Max plan have reported exhausting rolling windows startlingly fast. The grace window is more generous on paper than in practice.

To see why the paper-versus-practice gap is so wide, walk through the arithmetic. Suppose your weekly limit is worth a nominal budget of Claude usage, and Fable can consume up to half of it. Because Fable draws down that budget at twice the rate of Opus, the half you are allowed to spend on Fable buys you only about a quarter as much actual work as the same nominal budget would on Opus. A long agentic coding run that reasons deeply and emits large diffs can therefore eat a surprising fraction of a weekly allowance in a single sitting, which is exactly what heavy Max subscribers reported in the first days after relaunch. The 50 percent cap is not the real constraint. The burn rate inside it is. Two mechanics compound this: Claude enforces both a short rolling window refreshed every few hours and a separate weekly ceiling, so you can be blocked mid-task by the short window even while the weekly budget still has room. Once you cross the included Fable allowance you do not lose the model, you simply start paying metered usage credits at the same $10 and $50 rates as the API, which is fine for occasional use and painful for a workload that lives on Fable. The design nudges you, deliberately, to reserve Fable for the work that genuinely needs it.

Enterprise deserves its own sentence because the tiers behave differently. Standard Enterprise seats get no included Fable allowance and pay via usage credits from day one, while premium or consumption-based seats get the grace window - Digital Applied. And crucially, the Claude API and consumption-based Enterprise plans had full, uninterrupted access to Fable 5 throughout, subject only to the pause, because they were always metered rather than bundled - Anthropic. This is the quiet lesson of the plan layer: if you want Fable 5 with the fewest availability surprises, the pay-as-you-go API is the most predictable door, because it never depended on a promotional window that could expire or be clawed back. Subscribers got a genuinely good deal during the grace period, but they also got fewer free days than originally promised once the pause reset the calendar - TweakTown.

7. Availability by task: the safety classifiers and the Opus 4.8 fallback

This is the layer the user actually asked about most sharply, and it is the one almost every casual explainer gets wrong. Fable 5's availability is not just a matter of who you are and what you pay. It is a matter of what you ask. The model ships with safety classifiers that can decline specific requests, and when they fire, the request does not fail, it gets handed to a different model. This is the defining behavioral difference between Fable and its classifier-free twin Mythos, and it is the reason the same account can get Fable on one prompt and Opus on the next - Claude Platform Docs.

The domains that trigger a fallback are specific and, importantly, dual-use. Anthropic's own help documentation describes four areas where Fable 5 hands off, and the developer cookbook exposes three machine-readable categories that map to them. Understanding the list tells you exactly which of your tasks will and will not be answered by Fable itself:

• Offensive cybersecurity, meaning building exploits, malware, or attack tooling, tagged cyber in the API.
• The majority of biology, chemistry, and life-sciences queries, especially lab methods and molecular mechanisms, tagged bio.
• Model distillation and reasoning extraction, including attempts to pull out Fable's summarized thinking, tagged reasoning_extraction.
• A narrow set of frontier machine-learning tasks, such as distributed training infrastructure and accelerator design.

Now the mechanics, because the details determine how this feels in practice. When a classifier declines, the Messages API returns stop_reason: "refusal" as a successful HTTP 200 response, not an error, along with the category that triggered it - Claude Platform Docs. Your integration has to handle that as a normal, expected outcome rather than a failure. From there, the request can be retried on another Claude model, and Anthropic's designated fallback target is Claude Opus 4.8. On the API you can retry three ways: a server-side fallbacks parameter (in beta on the Claude API and Claude Platform on AWS only), client-side SDK middleware, or a manual retry you build yourself - Claude Cookbook. On Claude.ai the switch is automatic, and this is the part people miss: it is not silent. You see a notice explaining that the model switched, and the response is labeled with the model that actually answered - Anthropic Help Center. The frequency is low by design: fallbacks fire in fewer than 5 percent of sessions, so more than 95 percent of Fable conversations never touch a classifier at all.

The billing around this is unusually fair, and it removes a common objection. You are not charged for a request that is refused before any output is generated - Claude Platform Docs. When the fallback fires and Opus 4.8 answers, you are billed at Opus rates, which are half of Fable's, and a "fallback credit" refunds the prompt-cache cost of switching so you never pay the cache tax twice. On consumer plans the Opus response simply counts toward your usage limit at Opus rates. The net effect is that a blocked task costs you less, not more, which is a deliberate signal that Anthropic wants you to leave the safety on rather than route around it.

The dual-use tension is where this design gets genuinely hard, and it is worth being honest about the friction. A legitimate penetration tester asking Fable to explain an exploitation technique, or a graduate biologist asking about a molecular mechanism, is asking a question that is beneficial in their hands and dangerous in someone else's, and the classifier cannot read intent. It reads the topic. So both users will often find their requests routed to Opus 4.8, a capable model but a step down for exactly the frontier work they came for. The relaunch classifier, tuned to block the Amazon technique in over 99 percent of cases, also raised the false-positive rate, meaning some benign debugging and coding requests now trip the cyber filter that would have passed before June - MarkTechPost. If your work lives near these domains, expect more fallbacks than the topline 5 percent figure suggests. Architecting around this is straightforward once you accept it as the normal case: detect the fallback by inspecting the response, decide up front whether you want the automatic Opus answer or a hard stop, wire the fallbacks parameter or your own retry accordingly, and surface the switch to your users the way Claude.ai does. For a security or life-sciences product, the fallback is not an edge case to log and ignore, it is a core path to design deliberately.

One more piece of context frames why any of this exists: Anthropic classified Fable 5 and Mythos 5 at ASL-3 under its Responsible Scaling Policy, treating them as models with meaningful chemical, biological, radiological, and nuclear uplift risk - Interconnects. ASL-3 is what mandates the classifiers, the exemption process for vetted users, the bug bounty, and the rapid jailbreak response. So the task-level availability limits are not a product manager's preference, they are a formal safety commitment. If you are building anything that legitimately touches security research or the life sciences, plan your architecture around the Opus 4.8 fallback from the start rather than treating it as an edge case, because for your workload it will be the common case. Our deep dive on the Claude Agent SDK covers the retry and fallback plumbing in practical detail.

8. Mythos 5 and Project Glasswing: the version almost nobody can access

Any honest answer to "is Claude Fable available" has to address the model sitting right behind it, because for a specific and growing set of organizations, the real question is whether they can get Mythos 5, the unfiltered version. The answer for almost everyone is no, and that no is structural rather than temporary. Mythos 5 is not generally available. It is limited-release, offered only to approved customers through Project Glasswing - Claude Platform Docs. You cannot buy it, sign up for it, or hit it on a public endpoint. You get access by being vetted into a program, and the front door is your Anthropic, AWS, or Google Cloud account team, not a pricing page.

Project Glasswing is the reason Anthropic can justify building a Mythos-class model at all, and its shape reveals the company's actual customer for frontier capability. Glasswing is a collaboration oriented toward cyberdefenders, critical-infrastructure providers, and government-adjacent partners, with the express purpose of putting the strongest available model in the hands of the people defending against threats rather than the general public - Anthropic. The program has grown quickly, from roughly 50 partners in the spring to around 150 organizations across more than 15 countries by early June. During the export-control episode, it was Mythos access for a set of US organizations that was restored first, on June 26, four days before the general Fable relaunch, which tells you exactly whose availability Anthropic and the government prioritize when the stakes are high.

The distinction between the two models is the cleanest illustration of Anthropic's entire safety-and-availability philosophy. Fable and Mythos are the same underlying model; the only difference is that Fable carries the classifiers that route cyber, bio, and distillation requests to Opus 4.8, and Mythos does not - Anthropic. Anthropic has been candid that the reason it has not released a Mythos-class model publicly is that no one has yet developed safeguards strong enough to prevent misuse of the raw capability. Fable is the compromise: the public gets the intelligence, but not the parts of it that Anthropic judges too dangerous to leave ungated. That is why, for the overwhelming majority of readers, the practical answer is that Fable 5 is your ceiling, and it is a very high ceiling. For the history of how this two-track approach evolved, our Claude Mythos preview insider guide traces the lineage from the earlier preview through to today's split.

The strategic reading here, from first principles, is that Anthropic has effectively created a capability tier that maps to trust rather than to money. Normally you buy your way up a model ladder: pay more, get more. With Mythos, no amount of money moves you up, because the gate is vetting, not price. That is a genuinely new shape for AI availability, and it is likely a preview of how the most capable models will be distributed going forward: a public, filtered tier for everyone, and an unfiltered tier reserved for parties a government trusts.

For the small number of organizations that genuinely need the unfiltered model, the practical path is worth stating plainly, because it looks nothing like normal procurement. You do not sign up, you get sponsored: access to Mythos 5 runs through an account team and a vetting process oriented around who you are and what you defend, not what you are willing to spend. Cyberdefenders, critical-infrastructure operators, and national-security-adjacent research groups are the intended members, and the program's growth to roughly 150 organizations across more than fifteen countries suggests Anthropic is scaling it as fast as its trust process allows - Anthropic. For everyone else, the honest planning assumption is that Mythos is not on the table and will not be, so any architecture that quietly depends on unfiltered cyber or biology capability is a dead end on the public track. Build for Fable's guardrails, or move that specific workload to a model you host yourself.

Whether you find that reassuring or concerning, it is the reality Fable 5's availability sits inside, and it is not going away.

9. What Fable 5 actually costs, and the pricing myth to ignore

Cost is a form of availability, because a model you cannot afford to run at scale is not really available to you for that workload, and Fable 5 is expensive on purpose. The authoritative price, straight from Anthropic's pricing page, is $10 per million input tokens and $50 per million output tokens - Claude Platform Docs. That is exactly double Claude Opus 4.8, which sits at $5 and $25, and it is the single most important number for anyone deciding whether to build on Fable or route selectively to it. There is no introductory discount on the API. What you see is what you pay, from the first token.

That last point needs emphasis because there is a persistent myth circulating that Fable 5 is "$2 in, $10 out, rising to $3 and $15 in September." That pricing belongs to Claude Sonnet 5, a different and much cheaper model, not to Fable 5. The confusion is understandable, since both appear on the same pricing table, but conflating them will throw your cost model off by a factor of five. Sonnet 5's introductory rate of $2 and $10 runs through August 31 before stepping up to $3 and $15 - Claude Platform Docs. Fable 5 has never had a promotional rate. If a calculator or blog quotes you $2 for Fable input, it has mixed up the models, and you should discard the number.

The chart below puts the Anthropic ladder in perspective, and the shape is the whole point: Fable sits at the top by a wide margin, which is why fallback-to-Opus is not just a safety feature but a cost feature.

The headline rate is only the starting point, and the modifiers are where a real bill is won or lost. Prompt caching is the biggest lever: a cache read costs just $1 per million tokens, a tenth of the base input price, and for the long, repeated context that agentic work involves, this is the difference between a viable and a ruinous bill - Claude Platform Docs. The Batch API halves everything to $5 and $25 for non-urgent work. There is no long-context surcharge, so a 900,000-token request bills at the same per-token rate as a small one, which is unusual and generous. Two smaller factors round out the picture:

• US-only inference via the data-residency setting adds a 1.1x multiplier across all token categories.
• The new tokenizer used by Fable produces roughly 30 percent more tokens for the same text than older Claude models, quietly raising effective cost.

That tokenizer detail is the sort of thing that never appears in a pricing headline but shows up on your invoice, so budget for it. The practical takeaway on cost is that Fable 5 rewards discipline: cache aggressively, batch what you can, and route everything that does not genuinely need frontier reasoning to Opus, Sonnet, or Haiku. Teams that treat Fable as a default rather than a scalpel are the ones who post horror-story bills.

A concrete example clarifies the stakes. Imagine an agentic session that reads a large codebase into context, reasons over it, and emits a substantial patch: say 200,000 input tokens and 30,000 output tokens. At Fable's list price that single run costs roughly two dollars for input and a dollar and a half for output, about three and a half dollars before any caching. Run that a few hundred times a day across a team and you are into four figures of daily spend on one model. Now apply prompt caching to the static parts of that context, the code and instructions that do not change between runs, and the cached reads drop to a tenth of the input price, cutting the dominant cost line dramatically. The same workload on Opus 4.8 halves the remaining figure again. This is why routing discipline is not a nice-to-have: the gap between a cached, well-routed Fable workload and a naive one is the difference between a sustainable unit economic and a bill that ends the project. For a systematic approach, our guide to cutting LLM costs and the deeper true cost of LLM inference analysis both apply directly to Fable-era economics.

10. Why availability matters: capability, benchmarks, and lock-in

It would be easy to treat all of this friction as a reason to skip Fable 5, and for many workloads it is. But the friction only exists because the capability is real, and understanding how real is the difference between routing to Fable wisely and avoiding it out of hangover from the shutdown. Anthropic calls Fable 5 its most capable widely released model, and the benchmarks back a meaningful gap over the previous flagship rather than a marginal one - Anthropic. The clearest illustration is agentic coding, where Fable does not just edge out competitors, it separates from them.

On SWE-bench Pro, which measures resolving real software engineering tasks, Fable 5 scores around 80.3 percent, against roughly 69 percent for Opus 4.8 and well under 60 percent for the current OpenAI and Google flagships - Vellum. The chart below shows the spread, and the spread is the argument for tolerating Fable's availability quirks: for the hardest long-horizon work, nothing else public is close.

The pattern repeats across other frontier tests. On FrontierCode Diamond, a harder coding evaluation, Fable roughly doubles Opus 4.8, and on multimodal and vision tasks it posts averages well above the previous generation - Vellum. General reasoning benchmarks like GPQA Diamond are noisier and more contested, with different aggregators putting Fable, Opus, the current Gemini, and the current GPT within a few points of each other, so it would be dishonest to claim Fable dominates everywhere. Where it clearly leads is exactly the territory it was built for: long, tool-heavy, autonomous work. Anthropic's own benchmark visual makes the coding and agentic lead the centerpiece of its pitch.

Two caveats keep this honest. First, the coding lead is real but the general-reasoning picture is closer than the marketing implies. On knowledge benchmarks like GPQA Diamond, different aggregators put Fable, Opus 4.8, the current Gemini, and the current GPT within a few points of one another, and at least one ranks a competitor ahead, so anyone claiming Fable wins every category is overselling. The defensible claim is that Fable leads decisively on long-horizon agentic and software tasks and is competitive, not dominant, on general knowledge. Second, capability and safety were measured together: Anthropic reports that its bug bounty found no universal jailbreak across more than a thousand hours of red-teaming before launch, which is part of why the single working bypass from Amazon carried such weight. A model marketed as this robust cannot afford even one credible break, which is precisely why one triggered a government response. The takeaway for a builder is to match the model to the task rather than to the headline: if your workload is autonomous coding or multi-step research, Fable's lead is worth the friction and the price; if it is single-turn answering, classification, or retrieval lookups, the frontier gap narrows to where a cheaper model delivers the same user-visible quality at a fraction of the cost and none of the availability risk.

The first-principles reason availability matters so much for a model this capable is lock-in risk, and the June shutdown proved it is not hypothetical. If you rearchitect a product around Fable's specific strengths, its 1 million token context, its always-on thinking, its agentic memory, you are also inheriting its availability profile: a model a government can switch off in ninety minutes, that draws down your quota at double speed, and that refuses a defined set of tasks. The capability is worth designing for, but only if you design for its absence at the same time. The teams that came through the 19-day outage best were the ones whose systems could fall back to another model without a rewrite. That is not a reason to avoid Fable. It is a reason to abstract it. Our companion pieces on building apps with Fable 5 and Fable 5 for coding and company building go deeper on where that capability pays off in practice.

11. When Fable 5 is not your answer: the alternatives, scored

Given everything above, there are four clean situations where Fable 5 is not the model you should reach for: you are outside a supported region, you are on the free tier or out of quota, your task lives in a domain the classifiers block, or the price simply does not fit the job. In every one of those cases you need a substitute, and the good news is that the substitutes are excellent, starting with the model Anthropic itself falls back to. This section ranks the realistic alternatives on the dimensions that matter when you have lost Fable access: how easily you can actually get the model, how capable it is, what it costs, and how well it stands in for Fable's specific strengths.

The scoring below weights availability and access at 30 percent, raw capability at 30 percent, cost at 25 percent, and Fable substitutability at 15 percent, because when your first choice is blocked, getting a capable model you can actually run matters more than squeezing out the last point of benchmark performance. Each cell carries the number that justifies the score, and the table is sorted by final score from highest to lowest.

The clear takeaway is that the best Fable substitute is the model Anthropic already points to: Claude Opus 4.8 - Claude Platform Docs. It is a literal drop-in, using the same API, the same 1 million token context, and the same tools, at half the price, with none of Fable's classifier gating, since Opus is what those classifiers route to in the first place. If your only problem with Fable is cost or a blocked task domain, Opus 4.8 solves it without a single line of integration change. For the vast majority of "Fable is not available to me" situations, the answer is simply "use Opus 4.8," and it is not a downgrade so much as a different point on the same curve. Our Opus 4.8 benchmark and cost guide covers exactly where the two diverge.

If your problem is geography or a hard need to avoid any US export exposure, the calculus shifts toward open-weight models you host yourself, and here the field is genuinely strong in 2026. DeepSeek V4-Pro posts near-frontier coding numbers at a fraction of the cost and, because you run the weights, sidesteps the entire question of whether a government can revoke your access - OpenAI provides the counterpoint that a different closed provider like GPT-5.5 diversifies vendor risk without the self-hosting burden. The open-weight route trades convenience and, in some cases, raw capability for control and resilience, which is precisely the trade you want to make if the lesson you took from June is "never let one company's compliance department be a single point of failure." Our guides to DeepSeek V4, Kimi K2, and GLM 5.2 each go deep on the leading open-weight options.

Finally, if your problem is quota or cost on consumer plans, the honest move is to route by task rather than defaulting to Fable. Send the hard, long-horizon, agentic work to Fable within your allowance, drop to Sonnet 5 or Opus 4.8 for the middle of the distribution, and let Haiku 4.5 handle the high-volume, low-complexity tail at $1 and $5 - Claude Platform Docs. This is not a compromise, it is how experienced teams have always used model ladders, and Fable's double-speed burn rate makes the discipline more valuable than ever. The GPT-5.5 complete guide and the Gemini 3.1 Pro guide round out the cross-provider options for anyone who wants to hedge outside the Anthropic family entirely.

12. How to build so a single model outage cannot stop you

The deepest lesson of the Fable 5 saga is not about one model, it is about a structural fact that the June shutdown surfaced for the first time: the most capable models are now controlled goods, and their availability is a political variable, not just a technical one. For years the implicit assumption in building on top of an API was that the endpoint would be there tomorrow. June 12 killed that assumption. A frontier model went from generally available to globally offline in ninety minutes, and no amount of paying, coding, or planning on the customer side could have prevented it. If you build as though your chosen model is a permanent fixture, you have built a business with a dependency you do not control and cannot insure.

The correct response, reasoning from first principles, is not to avoid frontier models but to treat the specific model as a swappable component rather than a foundation. The value you deliver to a user is an outcome: a document written, code shipped, a research question answered, a company operated. The model is one input to that outcome, and inputs should be interchangeable. Concretely, that means designing for the fallback rather than bolting it on, and it means the abstraction living in your own system, not in a single provider's SDK. The teams that sailed through the 19-day outage were the ones whose architecture could point at Opus 4.8, or GPT-5.5, or a self-hosted open-weight model, without a rewrite, because they had never hard-coded a single model id into the heart of their product.

This is where the model-agnostic platforms earn their keep, and it is worth being concrete about what "agnostic" buys you. A well-designed agent platform separates the description of the work from the engine that does it, so the same instruction runs on whichever capable model is available and affordable at that moment. Platforms such as O-mega take this to its natural conclusion: you describe the autonomous company you want, the website, the app, the billing, the content, the operations, and a cloud agent workforce builds and runs it, drawing on whichever frontier model fits the task rather than on one hard-wired name. The point is not the branding, it is the architecture: when the model is an implementation detail behind an outcome you own, a government pulling one model offline becomes an inconvenience your system routes around, not an outage that stops your business.

There is a governance dimension to this too, and it is only going to grow. Fable 5's mandatory 30-day retention, its ASL-3 classification, and its export-control exposure mean that for regulated industries, the compliance surface of your AI stack is now part of your product risk. If you operate in finance, healthcare, or the public sector, the question is no longer only "which model is best" but "which model can I lawfully and reliably keep running," and the answer may be a boring, stable, generally available model rather than the newest frontier one. Yuma Heymans (@yumahey), who builds autonomous recruiting agents at HeroHunt.ai and writes throughout 2026 about the shift from AI tools to AI workforces, has made the point repeatedly that the durable advantage is not access to the smartest model on any given Tuesday, it is an operation that keeps delivering when the smartest model is unavailable. The Fable episode is the clearest evidence yet that he is right.

So build for the availability matrix, not against it. Use Fable 5 where its lead is decisive, long-horizon agentic and coding work inside your quota and outside the blocked domains. Fall back to Opus 4.8 by default, keep a cross-provider or open-weight option warm for the geographic and export edge cases, and put the routing logic somewhere you control. Do that, and the honest answer to "is Claude Fable available" stops being something you have to check anxiously each morning, because whether or not Fable is up, your system still works. For the broader strategic picture of how Anthropic reached this point, our look inside Anthropic's $965B IPO explains the commercial pressures pushing all of this, and the Sonnet 5 practical guide covers the workhorse most teams will actually run day to day.

Conclusion: a decision framework

Strip away the drama and the answer to the title is straightforward. Yes, Claude Fable 5 is available as of July 1, 2026, globally, on paid Claude plans and the developer API, for almost any task. The exceptions are specific and knowable: you cannot use it on the free tier, you cannot use it in sanctioned regions, you cannot use it under a zero-data-retention agreement, and you cannot use it for offensive cyber, most bio and chemistry, or distillation, where it will hand your request to Opus 4.8 and tell you it did. Everything else is a matter of quota and cost, both of which you control.

The decision framework that follows from the whole guide is compact. If you need the best public model for hard, long, agentic work and you are inside your allowance, use Fable 5, cache aggressively, and route the rest of your traffic elsewhere. If cost or a blocked domain is your problem, use Opus 4.8, which is the same shape at half the price with no classifier gating. If geography or export resilience is your problem, host an open-weight model and stop depending on any single vendor's compliance posture. And whatever you choose, put the model behind an abstraction you own, because the one durable lesson of June 2026 is that frontier availability is now conditional for everyone.

Fable 5 is a remarkable model wrapped in the most elaborate availability regime the industry has yet produced, and both facts are true at once because they are the same fact. It is gated precisely because it is powerful. Plan for the gates, build for their absence, and Fable 5 becomes exactly what it should be: a very sharp tool you reach for deliberately, not a foundation you pray stays online.

This guide reflects the Claude Fable 5 availability landscape as of July 2, 2026. Model access, pricing, quotas, and export-control status in this space change rapidly, and the June shutdown proved they can change within hours, so verify the current details with Anthropic before making commitments.