Claude Fable 5 and Mythos 5

What happened

On 9 June Anthropic launched two models at once:

• Claude Fable 5, which the company described as a Mythos-class model “that we’ve made safe for general use”. It said Fable 5 was “available everywhere today”.
• Claude Mythos 5, the same underlying model with safeguards lifted in some areas. It went first to partners in Project Glasswing as an upgrade to April’s Mythos Preview. Anthropic says it has “the strongest cybersecurity capabilities of any model in the world”.

Both cost $10 per million input tokens and $50 per million output tokens, less than half the price of Mythos Preview.

How capable is it?

Anthropic says Fable 5 leads in software engineering, knowledge work, vision and scientific research, and that the longer and more complex the task, the bigger its lead. The chart shows SWE-Bench Pro, a widely cited test in which models fix real issues in open-source projects:

• Claude Mythos 5 / Fable 580.3%
• Claude Mythos Preview77.8%
• Claude Opus 4.869.2%
• GPT-5.558.6%
• Gemini 3.1 Pro54.2%

The table shows the higher of the Mythos 5 and Fable 5 scores; Anthropic says they are within 1–3 points on this test. Source: Anthropic, “Claude Fable 5 and Claude Mythos 5” benchmark table, 9 Jun 2026

On GDPval-AA, a test of real knowledge work, it scored 1932, ahead of Opus 4.8 (1890), GPT-5.5 (1769) and Gemini 3.1 Pro (1314).

It did not lead everywhere. On OSWorld-Verified, which tests using a computer, Mythos Preview scored slightly higher (85.4% against 85.0%), and it also edged ahead on Humanity’s Last Exam with tools (64.7% against 64.5%).

Anthropic and early customers gave concrete examples, none independently checked:

• Big code migrations. Stripe said Fable 5 migrated an entire 50-million-line Ruby codebase in a day, work that would have taken a team more than two months by hand.
• Playing games by sight. Earlier Claude models struggled with Pokémon FireRed even with helper tools; Fable 5 finished it using only screenshots.
• Taking notes over long tasks. In the card game Slay the Spire, giving it a file to keep persistent notes improved its results three times as much as it did for Opus 4.8.
• Science. Anthropic’s protein-design experts used Mythos 5 to speed up parts of drug design by about ten times; 9 of 14 protein targets produced strong candidates. In blind comparisons, the company’s scientists preferred Mythos’s molecular-biology hypotheses about 80% of the time.

How the safeguards work

Anthropic says Mythos-class models “have reached a threshold where they present significant risks”. Fable 5’s main protection is a set of classifiers: separate, smaller AI systems that watch for possible misuse.

1. 1Classifier checkThe request first passes through separate safety classifiers looking for cybersecurity, biology and chemistry, or “distillation” (mass extraction of the model’s abilities to train another model).
2. 2Not flaggedMost conversations. Fable 5 answers normally. Anthropic says over 95% of sessions never trigger a fallback, and there Fable 5 performs essentially like Mythos 5.
3. 3FlaggedClaude Opus 4.8, the next model down, answers instead and the user is told. Anthropic says this beats a flat refusal.
4. 4Monitoring afterwardsTraffic on Mythos-class models is kept for 30 days to spot new jailbreaks, then deleted in almost all cases.

Source: Anthropic announcement, 9 Jun 2026

Anthropic’s test results: an outside bug bounty produced no universal jailbreak in over 1,000 hours of testing, and outside red-teaming groups had not found one on long agentic tasks, though the UK AI Security Institute (UK AISI) “made progress towards one” in a brief initial testing window. The company says completely preventing universal jailbreaks is “likely impossible”; the aim is to make them slow and costly enough to catch before they are used at scale.

To ship quickly and safely, Anthropic tuned the safeguards to be cautious. It acknowledged that harmless requests would sometimes be caught, and most biology and chemistry questions go to Opus 4.8 for now.

How it compares

In Anthropic’s table, Fable 5 / Mythos 5 leads OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Pro on most listed tests. But the starred rows (cybersecurity, biology, Terminal-Bench 2.1 and others) show a bigger gap between Mythos 5 and Fable 5: because the safeguards hand those tasks to Opus 4.8, Fable 5 performs closer to Opus 4.8 there than the table suggests.

In Anthropic’s automated alignment assessment, Mythos 5’s level of misbehaviour, such as deception or cooperating with misuse, was low and similar to Opus 4.8’s.

Things to keep in mind

• The scores come from the company. Anthropic published the table, and it shows only the higher of the Mythos 5 and Fable 5 scores.
• Expect some answers from a smaller model. Sensitive topics go to Opus 4.8, and Anthropic itself says the safeguards are “stricter than would be ideal”.
• A new data-retention rule. Business traffic on Mythos-class models must be kept for 30 days. Anthropic says it won’t train on this data, but it is a real cost for some customers.
• Trouble came fast. Three days after launch, the US Commerce Department imposed export controls over a reported jailbreak, and Anthropic had to switch both models off. See US export order forces Anthropic to suspend its newest models.

How to try it

• Subscribers. At launch, Fable 5 was included at no extra cost on Pro, Max, Team and seat-based Enterprise plans until 22 June, after which it required usage credits. When it returned on 1 July, Pro, Max, Team and select Enterprise plans could use it for up to 50% of weekly usage limits until 7 July, then via usage credits. For plan prices, see our Claude product page.
• Developers. Through the Claude API as claude-fable-5, at $10 per million input tokens and $50 per million output tokens.
• Later versions. Fable 5.1 (1 September) cut costs and loosened some safeguards; Anthropic says Opus 5.5 (22 September) performs at Fable 5.1’s level on most work.