Claude Fable 5.2 vs. Mythos 5.2: What’s the Difference?
Fable 5.2 and Mythos 5.2 are the same model with different safeguards. Here’s who can access Mythos and what the benchmarks show.
👁 115 reads
Imagine asking Claude Fable 5.2 to write an exploit for a vulnerability you just found on your own server. You get an answer, but not from Fable. A classifier deems the request too risky and routes it to the older Claude Opus 4.8. The same model, without that restriction, is available to security organizations under the name Mythos 5.2. Anthropic says that in an October run, it reproduced 89.6% of 1,829 known vulnerabilities.
That’s the main difference between two models announced on November 9, 2026. I read the announcement, system card, documentation for both models and the rules for the program that provides Mythos access. The takeaway is straightforward, if disappointing: these aren’t two models you get to choose between. They’re one model with two sets of rules, and Anthropic decides which set you get.
Same model, different safeguards
The system card describes them as two configurations of the same new model. The Mythos 5.2 documentation even has the subtitle “Claude Fable 5.2 for verified organizations.” Both have a 1-million-token context window, a maximum output of 128,000 tokens, knowledge through June 2026 and identical pricing: $8 per million input tokens and $40 per million output tokens. Their API identifiers differ: claude-fable-5-2 and claude-mythos-5-2.
The difference is in the classifiers placed in front of Fable 5.2. They block high-risk cybersecurity tasks, including penetration testing, exploit writing and finding vulnerabilities in binary files. They also block most biology research tasks. A blocked request isn’t refused outright; it’s routed to a less capable model: Opus 4.8 for cybersecurity and Opus 5 for biology. The system card adds that Fable 5.2’s cybersecurity safeguards enforce the same policy as Opus 5.5.
This mechanism has a side effect Anthropic itself describes. In prompt-injection tests, Fable 5.2 switches to Opus 4.8 more often than previous models, making attacks more likely to succeed. When it responds on its own, the system card says it’s among the company’s most robust models. The weak link, then, isn’t Fable but the fallback model that replaces it.
Fewer false alarms and one deliberate restriction
When Anthropic split Fable and Mythos in version 5.1, the most common complaints were about overzealous filters. Anthropic says Fable 5.2 has about 55% fewer false blocks than Fable 5.1, with about 60% fewer in cybersecurity and 40% fewer in biology. These are company figures, and there’s no independent measurement yet, but the direction is clear: an administrator asking about firewall configuration should be less likely to be routed to Opus 4.8.
There’s also a new, deliberate restriction. According to the announcement, Anthropic trained the model to reduce its abilities in two narrow areas: kernel optimization for accelerators and improving the architecture of language models. In those areas, it’s intended to perform at the level of Opus 5.5. Agentic coding, hardware design and machine-learning research were left untouched. Some topics, such as kernels for custom accelerators, are routed by Fable to Opus 5. Anthropic doesn’t want its strongest publicly available model helping competitors build more AI.
What Mythos 5.2 can do without the safeguards
The figures in Anthropic’s announcement compare Mythos 5.2 with Mythos 5.1 and with Opus 5.5 running without safeguards:
| Test | Mythos 5.2 | Mythos 5.1 | Opus 5.5 without safeguards | GPT-6 Cyber |
|---|---|---|---|---|
| CyScenarioBench (cyber scenarios) | 76.3% | 61.7% | 67.6% | 61% |
| CVE Rediscovery (reproducing known vulnerabilities) | 89.6% | 78.3% | 81.3% | - |
| De novo binder design, 24 h (protein design) | 85.1% | 79.1% | 82.6% | - |
| Terminal-Bench-Science 0.1 | 75.3% | - | 58.7% | - |
For comparison, Fable 5.2 scores 74.6% on Terminal-Bench-Science, nearly matching Mythos. Where the safeguards don’t apply, the two models are practically indistinguishable. GPT-6 Cyber’s CyScenarioBench result comes from a different set of tasks, since each lab gets its own tasks for this benchmark, so the comparison is indicative only.
Anthropic’s announcement also includes reports from its partners. One security company said that every other model it tested missed one in six confirmed bugs found by Mythos 5.2. Mozilla said Mythos 5.2 roughly doubled vulnerability detection rates compared with Mythos 5.1. In biology, Mythos 5.2 correctly modeled the structures of 28 of 34 AAV virus capsids it worked on with Dyno Therapeutics, Columbia and Biohub. The system card notes that the model didn’t exceed the CB-2 threshold in biology and receives CB-1-level protections, as many previous models do.
Testing notes
I couldn’t test Mythos 5.2 because I don’t have access to it: it isn’t available on OpenRouter, and Anthropic limits access to organizations in its Cyber Verification Program and Life Sciences Verification Program. All Mythos figures in this article therefore come from Anthropic’s announcement and system card. I tested Fable 5.2 myself on 27 tasks. It scored 100% on coding, puzzles and image questions, matching Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon. The test was conducted in Polish, not English.
Who can get Mythos 5.2?
There’s no “buy” button for Mythos 5.2. The documentation says it’s available only to organizations verified through Anthropic’s programs. There are two: the Cyber Verification Program for security teams and the Life Sciences Verification Program for life-sciences researchers, which the company recently expanded. A third route is Claude Security, a code-scanning tool for Claude Enterprise customers. According to the system card, it runs on Mythos 5.2 and is available to all Enterprise customers, including those outside the verification program.
The Cyber Verification Program has three access levels: Defense, Red Team and Specialized. An organization submits one application through the Verification Portal, and Anthropic promises a decision within seven business days. Individuals, including independent researchers, open-source project maintainers and bug bounty hunters, can apply only for Defense and must have a paid plan. Specialized access requires a detailed review involving the US government. Data retention is mandatory, and Defense-level users must use phishing-resistant authentication and stop using API keys by December 15, 2026.
The program rules don’t list eligible countries. Anthropic says only that it’s working to expand access in the US and internationally, and that it considers the legal environment and who an applicant’s work serves. Organizations outside the US can apply, but approval isn’t guaranteed. Organizations working mainly for the military, intelligence services or police receive closer scrutiny. Check with Anthropic to confirm your organization’s eligibility. For most users, Fable is the available option; Mythos access is restricted to approved organizations and Claude Enterprise customers using Claude Security.
Which one should you choose?
In practice, there’s one answer to the question in the title. If you don’t work professionally in offensive security or biology, Fable 5.2 is the option available to you. For coding, analysis and working with documents, its safeguards shouldn’t get in your way. If you run a security team, consider applying to the program: the vulnerability detection results are notable, and the API prices are identical. Anthropic’s announcement and system card provide more information about the models and their access programs.
Frequently asked questions
Is Claude Mythos 5.2 more capable than Fable 5.2?
Not in terms of the underlying model: they’re the same model. Mythos has an advantage only where classifiers block Fable, namely offensive cybersecurity and biology research. On other tasks, their results are practically the same, such as their 74.6% and 75.3% scores on Terminal-Bench-Science.
How much does Mythos 5.2 cost?
It costs the same as Fable 5.2: $8 per million input tokens and $40 per million output tokens.
What happens when Fable 5.2 blocks a request?
The request is routed to a less capable model: Opus 4.8 for cybersecurity and Opus 5 for biology. You still get an answer, but not from Fable.
Can an individual get access to Mythos 5.2?
Only in a limited capacity. Independent researchers and bug bounty hunters can apply to the Cyber Verification Program for Defense-level access if they have a paid Claude plan. Anthropic makes the decision.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Sources
- Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (November 9, 2026)
- Anthropic: System Card Claude Fable 5.2 & Claude Mythos 5.2 (November 9, 2026)
- Claude documentation: Claude Mythos 5.2 (checked November 9, 2026)
- Claude documentation: Claude Fable 5.2 (checked November 9, 2026)
- Claude Help Center: Cyber Verification Program (checked November 9, 2026)
- Claude: Life Sciences Verification Program
- Claude Security
