Claude Fable 5: Anthropic's First Public Mythos-Class Model Unleashed

Anthropic spent two months explaining that its most capable model wasn’t ready for a general release. The cybersecurity capabilities alone made the company hesitant, and they meant it: Claude Mythos Preview has been running inside a closed vetted-partner program since April, available only to approved cyber defenders and critical infrastructure providers. On June 9, Anthropic decided the safeguards were ready to hold.

Claude Fable 5 launched Tuesday as the first generally available Mythos-class model. Alongside it, Anthropic released Claude Mythos 5, an updated version of the restricted model for existing Project Glasswing partners. Both run on the same underlying model weights. What separates them is a set of safety classifiers that monitor incoming requests and reroute sensitive ones before the main model ever responds.

Fable and Mythos: Same Model, Different Gates

The names are deliberately paired with each other. Fable comes from the Latin fabula, the same root as the Greek mythos. Anthropic’s point is that the safety classifier layer is the only meaningful difference between the two models. Strip those classifiers away and you have Mythos. Leave them in place and you have Fable.

In practice, Fable 5’s classifiers cover three categories. Cybersecurity exploitation and offensive operations trigger a fallback. Biology and chemistry requests in dual-use territory trigger a fallback. And what Anthropic calls distillation, meaning large-scale systematic attempts to extract Fable’s capabilities to train competing models, triggers a fallback as well. When any of these happen, the response comes from Claude Opus 4.8 instead, and users are notified that the switch occurred. Anthropic’s early data shows more than 95% of Fable sessions see no fallback at all.

But that 5% has structure you’ll want to understand before building against it. In the domains the classifiers cover, the fallback rate isn’t 5%. Vals AI ran Fable 5 against their full benchmark suite and reported close to 100% refusal on biology and cybersecurity questions specifically. If your integration touches those areas, you’re paying Fable 5 prices and getting Opus 4.8 responses. For everything outside those domains, the classifier architecture stays genuinely invisible in ordinary use.

Project Glasswing and Mythos 5

Project Glasswing launched in April as Anthropic’s controlled deployment program for defensive cybersecurity work. The founding coalition included AWS, Apple, IBM, Microsoft, Cisco, CrowdStrike, the Linux Foundation, and others, with the stated goal of putting Mythos-class capability to work for defenders before that capability proliferates more widely. IBM named Glasswing explicitly in its May 2026 enterprise security announcements, describing how the program feeds into IBM Concert’s vulnerability-scanning platform and IBM’s autonomous security services.

By the June 9 launch, roughly 200 organizations had participated in Glasswing, and partners had collectively found more than 10,000 high- or critical-severity vulnerabilities across widely used software. Claude Mythos 5, which rolled out to all existing Glasswing partners on June 9, upgrades them from Mythos Preview at a dramatically lower cost: $10 per million input tokens and $50 per million output, compared with the $25 input and $125 output that Mythos Preview carried after Anthropic’s initial $100M credit commitment ran out.

Anthropic is also planning a separate trusted access tier for biology researchers, where the biology and chemistry classifiers are removed while cyber safeguards remain. That program hadn’t opened at launch, but the intent was to enroll select life science organizations in the weeks following release.

Benchmarks and Where the Model Stands

Fable 5 debuted at number one on Vals AI’s overall index, scoring 75.14% across their benchmark suite and ranking first of 29 models with the Opus 4.8 fallback enabled. It also launched at number one on Artificial Analysis’s Intelligence Index with a score of 64.9, approximately five points ahead of the nearest non-Anthropic competitor. On LLM-Stats’s SWE-bench Verified leaderboard, Fable 5 reached 95.0%, the highest published score for a generally available frontier model at launch.

For coding and complex reasoning, those numbers held up to independent verification. Vals AI noted that counting refusals as failures (rather than allowing the Opus 4.8 fallback to answer) collapsed Fable 5’s scores in flagged domains significantly. On GPQA Diamond, a set of graduate-level science questions, Fable 5 fell from 93.18% accuracy with fallback enabled to 55.56% when refusals were scored as failures outright. So the composite number depends heavily on how you count the classifier behavior.

Vision is also weaker than the broader benchmarks suggest. Fable 5 placed tenth in Roboflow’s vision evaluations, behind Gemini 3.5 Flash, Gemini 3.1 Pro, and GPT-5.4. Artificial Analysis’s independent testing confirmed the same vision gap. For coding and reasoning work, the rankings held and the gap over competitors was real. For vision-intensive tasks, the model was solidly mid-field.

The UK AI Security Institute and the 73% Figure

Pre-launch, Anthropic worked with the UK AI Security Institute on robustness testing for Fable 5’s classifiers. In a compressed testing window before release, UK AISI developed a jailbreak technique for single-turn cybersecurity queries within a few hours of starting. After roughly two days of testing, the team expanded the approach to multi-step agentic tool calls. Full long-form malicious agentic tasks weren’t completed in the time available, and UK AISI characterized their findings as interim results from a compressed window.

The number that circulated widely: Gina Neff, Professor of Responsible AI at Queen Mary University London, told the BBC that UK AISI testing found the model could exploit defenses and systems 73% of the time. That’s a real finding from real testing, and it’s worth taking seriously. It’s also describing performance in a jailbreak-assisted benchmark setting, not ordinary unprompted use against an average target. Anthropic said it would continue working with UK AISI on classifier robustness testing after launch. They didn’t claim the safeguards were airtight, only that universal jailbreaks were slow and costly enough to detect before being used at scale.

Pricing and Availability

Both Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output. That’s exactly double Claude Opus 4.8’s standard rate of $5 input and $25 output. Prompt caching brings cached input down to $1 per million tokens, a 90% input discount. Batch API pricing runs $5/$25, half the standard rate.

The API model ID for Fable 5 is claude-fable-5. At launch, Fable 5 was generally available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Mythos 5 requires approved Glasswing access and isn’t available through standard channels.

For subscription users, Anthropic offered an included window: through June 22, Fable 5 was part of Pro, Max, Team, and seat-based Enterprise plans at no extra cost. On June 23, Anthropic planned to shift Fable 5 to usage credits for those plans, with the intent to restore it as a standard subscription feature once capacity allowed. So the window was short even before anything else changed.

The Mandatory Data Retention Change

One policy shift that matters for enterprise users: Anthropic is requiring 30-day data retention on all Fable 5 and Mythos 5 traffic, across both first- and third-party surfaces. That requirement overrides any prior zero-retention agreements customers had in place for these model tiers.

Anthropic’s stated purpose for this is safety. The retained data helps the company detect novel jailbreak attempts and attacks that work by spreading behavior across many requests rather than a single interaction. The data isn’t used for model training, human access to it is fully logged, and deletion happens after 30 days in almost all cases. If your organization had previously negotiated zero-retention terms for API usage, those terms don’t extend to Fable 5 or Mythos 5, and you’ll want to verify compatibility with your compliance requirements before routing sensitive data through either model.

Update: Access Suspended June 12

Three days after launch, Fable 5 and Mythos 5 went offline for all customers.

On June 12, the US Commerce Department issued an export control directive requiring Anthropic to suspend access to both models by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. To ensure compliance, Anthropic disabled both models globally for all customers. Access to Claude Opus 4.8 and other Anthropic models was not affected.

Anthropic published a statement disagreeing with the basis for the directive. The company said it had reviewed the specific jailbreak technique cited by the government, found that the same approach produced comparable results on other publicly available frontier models including GPT-5.5, and characterized the situation as a misunderstanding that it was working to resolve. The subscription window through June 22 ended up being irrelevant: the models were offline before most users had time to form a real opinion about them.

For the full account of what happened and what it means for developers: U.S. Government Suspends Claude Fable 5, Declares It a Cyberweapon.

Update: Access Restored July 1

The outage was temporary. The US Commerce Department withdrew its directive on June 30, and Anthropic redeployed Fable 5 and Mythos 5 globally on July 1, 2026, ending an 18-day suspension. To satisfy the government, Anthropic shipped a new safety classifier that it says blocks the cited jailbreak in more than 99% of cases.

The subscription window described above never really took effect. Included access was extended to July 12 and then July 19 before Fable 5 moved to metered usage credits on July 20, at the same $10 per million input and $50 per million output. Max and Team Premium seats keep 50% included access; Pro and Team Standard pay per use.