Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026, describing them as two access profiles for the same underlying model. Fable is the broadly available version, with cybersecurity and biology safeguards. Mythos gives vetted organizations access for work in those dual-use fields.
The release targets demanding, open-ended technical projects: long-running coding agents, scientific research and engineering workflows. Anthropic claims faster output and substantial capability gains over the 5.1 versions. Its developer documentation also reveals changes that deserve attention before an upgrade, including always-on adaptive thinking, restrictions on forced tool calls and a 30-day data-retention requirement. (Announcement, model documentation)
One Model, Two Access Profiles
Fable and Mythos should be understood as one model release with different safeguards and access conditions, not two independently trained models competing on general intelligence.
Anthropic says Fable’s cybersecurity and biology protections allow it to offer the model broadly. Organizations pursuing beneficial cybersecurity or biomedical research can apply for Mythos through its Cyber Verification and Life Sciences programs. The documentation says Mythos shares Fable’s specifications and pricing; access requires verification or a conversation with an account team. (Anthropic announcement, model documentation)
That distinction matters when interpreting benchmark scores. A Mythos result on vulnerability discovery describes what Anthropic reports for the restricted access profile. It should not be treated as a promise that public Fable access will answer the same requests.
My view is that the access policy is a central product feature here. A security team deciding whether this release fits its workflow needs to assess both technical capability and eligibility. The shared model does not make the two access paths interchangeable.
Fable is listed as active on the Claude API, Amazon Bedrock, Google Cloud through both Google-managed and Anthropic-managed offerings, Microsoft Foundry, and Claude Platform on AWS. The Claude API model ID is claude-fable-5-2; Bedrock uses anthropic.claude-fable-5-2. (Platform documentation)
What Anthropic’s Benchmarks Actually Show
Anthropic’s published results put Fable 5.2 ahead of Fable 5.1 and Opus 5.5 on several coding, scientific and computer-use evaluations.
| Evaluation | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|
| FrontierCode 2.0, agentic coding | 49.1% | 32.4% | 45.5% |
| FrontierSWE v2, agentic coding | 69.9% | 56.3% | 62.3% |
| Terminal-Bench-Science 0.1 | 74.6% | 52.6% | 58.7% |
| RoboDojo Sim, robotics | 64.9% | 27.6% | 39.4% |
These are Anthropic-reported results, not independent measurements supplied with this article. The company says scores generally use each model’s highest available effort unless otherwise noted. FrontierCode lists Fable 5.2 and 5.1 at Xhigh effort, and RoboDojo Sim was run internally. Those conditions matter when translating a leaderboard result into expected cost or latency. (Benchmark table and notes)
The gains also vary considerably by task. On FrontierCode, Fable 5.2 leads Opus 5.5 by 3.6 percentage points. On Terminal-Bench-Science, the lead is 15.9 points. That supports a workload-specific upgrade decision: the published advantage is much larger on some evaluations than others.
A prominent Hacker News launch thread also carried claims that Opus 5.5 had been weakened ahead of Fable 5.2. One commenter who reran a small private evaluation reported results within the noise of prior runs, apart from a faster time to first token. That exchange is anecdotal and does not establish either a pre-release downgrade or broad user sentiment. (Hacker News discussion)
Cybersecurity Gains Need the Right Denominator
Anthropic reports a score of 89.6% for Mythos 5.2 on its internal CVE Rediscovery evaluation, compared with 78.3% for Mythos 5.1 and 81.3% for Opus 5.5 without safeguards. The company describes those results as finding 14.4% more vulnerabilities than Mythos 5.1 and 10.2% more than Opus 5.5. (Anthropic announcement)
The arithmetic explains the wording:
- Against Mythos 5.1: an 11.3-percentage-point increase, equivalent to about 14.4% relative improvement.
- Against Opus 5.5: an 8.3-percentage-point increase, equivalent to about 10.2% relative improvement.
That is a meaningful improvement within Anthropic’s test. A known-vulnerability rediscovery evaluation, however, measures a particular task. It cannot by itself tell a security team how many previously unknown vulnerabilities the model will find in its own codebase.
Anthropic’s CyScenarioBench comparison needs additional care: the launch page explicitly says each lab’s scores use unique tasks. I would avoid treating that table as a clean, like-for-like ranking between vendors.
Scientific Claims Depend on the Agent Setup
Anthropic says prerelease work included difficult software migrations, three working electronic devices and solutions to two open problems in black-hole physics. Those are company-reported outcomes, with different workflows behind them. (Launch announcement)
One setup detail worth close attention appears in The Next Web’s launch coverage: Anthropic used a team of 1,000 Fable 5.2 agents to formalize the proof of Fermat’s Last Theorem in 30 hours.
Formalization means translating a proof into a form that software can check step by step. It is different from discovering a new proof of the theorem. The scale of the agent team also matters: “30 hours” describes elapsed time for that coordinated setup, not a single chat session or a disclosed total compute budget.
For developers, the useful lesson is about orchestration. A model’s ability to contribute to a large parallel workflow can be valuable, but elapsed time alone leaves unanswered questions about token consumption, coordination overhead and reproducibility.
Specifications and Pricing Favor Selective Use
Fable 5.2 has a one-million-token context window and supports up to 128,000 output tokens. It accepts text and images and produces text. Anthropic lists June 2026 as both its reliable knowledge cutoff and training-data cutoff. (Model specifications)
The context window gives an agent room for large inputs and accumulated working context. The output limit is a separate ceiling; neither figure guarantees that every long task will be handled accurately or completed within one response.
Standard and fast-mode pricing are:
| Mode | Input per million tokens | Output per million tokens | Anthropic’s speed claim |
|---|---|---|---|
| Standard | $8 | $40 | Up to 30% faster output than the 5.1 versions |
| Fast | $16 | $80 | Up to 2.5 times faster token generation |
The speed claims concern output generation, so they should not be read as guaranteed reductions in total agent completion time. Tool execution, external services and other workflow steps can still dominate elapsed time. (Announcement and fast-mode pricing)
At standard rates, 100,000 billed input tokens plus 10,000 billed output tokens would cost $1.20. The same token counts in fast mode would cost $2.40. These are calculated examples using the published rates, without caching or batch discounts.
For comparison, the documented Opus 5.5 rates are $4 per million input tokens and $20 per million output tokens. That makes standard Fable twice as expensive for identical billed token counts. Actual task costs may differ because the models can use different amounts of reasoning and output. (Pricing comparison)
Anthropic itself recommends starting with Opus 5.5 for most coding and knowledge work, then moving to Fable or Mythos when complexity warrants it. Fable can also advise smaller models on difficult steps or orchestrate work delegated to them. (Model-selection guidance)
I’d treat that as the sensible deployment pattern: escalate hard tasks selectively, then measure whether the extra capability reduces retries or unfinished work enough to justify the higher token price.
Frequently Asked Questions
3 questions
1Are Fable 5.2 and Mythos 5.2 Different Models?
Anthropic says they share the same underlying model. Their safeguards and access conditions differ, particularly for cybersecurity and biology work.
