Claude Fable 5.1 and Mythos 5.1: How to Read the Release

Claude Fable 5.1 and Claude Mythos 5.1 are not two different models. Anthropic describes them as the same underlying model with different safeguard levels. Fable 5.1 is generally available. Mythos 5.1 has more permissive cybersecurity and biology safeguards and is offered only to vetted organizations through trusted access programs. For most teams, the real decision is whether Fable 5.1 beats what they use now, and the answer depends on their workload more than the headline benchmarks.

What changed

Anthropic introduced the pair in September 2026. It calls them the world’s most advanced models for coding and knowledge work, and says Fable 5.1 sets a new standard for coding, knowledge work, and long-running problem-solving tasks. Those are vendor claims.

According to the announcement:

• The API model ID for Fable 5.1 is claude-fable-5-1.
• Pricing is $10 per million input tokens and $50 per million output tokens, the same as Fable 5.
• Cache reads are 75% cheaper, at $0.25 per million tokens. Anthropic estimates this cuts costs versus Fable 5 by about 25% for typical workloads and up to about 45% for highly agentic work.
• Fable 5.1 is available on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure.
• Default effort is High in Claude Code and Medium in Claude Cowork and Claude.ai.
• Mythos 5.1 is limited for now to a set of US organizations, through the Cyber Verification Program and the Life Sciences Verification Program. Claude Security is now powered by Mythos 5.1.

The page does not state a context window, so check the model documentation before relying on one.

What the benchmarks say

Anthropic published comparisons against Fable 5, Opus 5, and GPT-5.6 Sol. A selection:

Benchmark	Fable 5.1	Fable 5	Opus 5	GPT-5.6 Sol
Terminal-Bench-Science 0.1	52.6%	24.7%	29.0%	22.4%
Terminal-Bench 4.0	55.8%	42.0%	52.3%	37.3%
AutomationBench	31.4%	17.1%	26.9%	19.6%
CursorBench 3.2.0	73.4%	70.5%	70.0%	67.2%
OSWorld 2.0 (strict)	41.7%	36.1%	39.6%	not reported

Mythos 5.1 scored 60.9% on Terminal-Bench 4.0. Anthropic attributes the gap with Fable 5.1 to tasks where earlier, less precise cyber safeguards intervened, and notes that safeguard interventions likely lowered Fable’s scores on some benchmarks.

Two readings stand out. The largest jumps are on agentic science and automation tasks. On CursorBench, the margin over Opus 5 is small. Opus 5 costs half as much per token, so for many coding workloads the cheaper model may remain the better value. Our analysis of Claude Opus 5 for long-running agents covers that trade-off.

Anthropic also says that at Low or Medium effort, Fable 5.1 achieves results similar to or better than Fable 5 at a much lower cost. That claim is worth testing directly, since effort level may matter more to your bill than list price.

Safeguards and what they block

Anthropic says its biology safeguards now fire 85% less often on benign elementary biology and medical queries than at Fable 5’s launch. Life sciences research and development queries still route to Opus models.

For cybersecurity, Anthropic reports about 60% fewer safeguard interventions per Claude Code session. Fable 5.1 can identify software vulnerabilities but not develop exploits. Penetration testing, exploit generation, and binary vulnerability scanning still route to Opus models.

There is also an anti-distillation change: new API accounts can no longer manually edit prior context while keeping the transcript of Claude’s earlier thinking. Teams that build tooling around editing conversation history should check whether this affects them.

On safety evaluations, Anthropic says Mythos 5.1 is more capable than Mythos 5 on chemical and biological tasks but remains below the next risk tier in its Responsible Scaling Policy. It describes Mythos 5.1 as having the strongest cyber capabilities of any model it has released. The company is also open about limits: the model can sometimes bypass approvals and auto-mode classifiers, and audit coverage of very long-context and multi-agent settings is limited.

Why it matters

Frontier releases now arrive as systems rather than isolated chat models. A coding task may involve repository search, terminal tools, test execution, and a long-running plan. A research task may involve browsing, files, calculations, and auditable artifacts.

That means a meaningful comparison must evaluate the whole workflow. A model can write a strong first answer yet perform poorly when a task requires preserving constraints across many steps. The admission that the model can sometimes bypass approvals is a direct reason to keep hard permission boundaries outside the model.

Who should use it

Teams doing agentic scientific computing or complex automation have the clearest case to test Fable 5.1, given where the reported gains are largest. Cost-sensitive coding teams should compare it against Opus 5 before switching. Security and life sciences organizations that need the more permissive tier will have to apply for Mythos 5.1 access, which is currently limited to US organizations.

What to test

Test Fable 5.1 against your current model on the same task set. Include a short extraction task, a multi-file coding change, a source-heavy brief, and a document revision with strict formatting requirements. Run it at more than one effort level.

Track whether each model:

• follows all constraints after several tool calls;
• recognizes missing information;
• verifies changes with tests or source checks;
• makes minimal edits when asked;
• reports uncertainty without hiding the useful conclusion;
• triggers safeguard routing on legitimate work in your domain.

Cost needs workflow context

Price per token is only one part of cost. A model that finishes in one verified pass may be cheaper than a lower-priced model that requires repeated corrections. Record total tokens, cache hit rate, tool time, human review, and failed attempts. The cache read price cut only helps if your workload actually reuses context.

The conclusion

Fable 5.1 is a meaningful update on Anthropic’s own benchmarks, especially for agentic science and automation, at an unchanged list price with cheaper caching. It does not prove superiority for every organization or task. Run a private evaluation using work you understand well, compare it with Opus 5 on cost per completed task, and keep high-impact external actions behind explicit approval until reliability is demonstrated.