Claude Fable 5.1 lands. The real news is the price cut.

Anthropic announced Claude Fable 5.1 and Claude Mythos 5.1 this week, the first point release for the Mythos-class tier it introduced in June 2026. The benchmark jumps are large, and we will get to them. But the detail that changes the decision for most teams is not a benchmark. It is the bill: Anthropic says typical workloads now cost 25% less to run, highly agentic workloads up to 45% less, and cache reads drop 75% to $0.25 per million tokens.

What actually shipped

Fable 5.1 and Mythos 5.1 are the same underlying model with different safeguard levels. Fable 5.1 is generally available under the model ID claude-fable-5-1; Mythos 5.1 carries reduced restrictions and is reserved for vetted cybersecurity and life sciences professionals through Anthropic's verification programs. List pricing stays at $10 per million input tokens and $50 per million output tokens, the same as Claude Fable 5. The savings come from efficiency: the model reaches answers with fewer tokens, and the cheaper cache reads compound on agentic workloads that re-read long contexts on every turn.

The benchmark picture

The headline numbers, all from Anthropic's own announcement:

Benchmark	Fable 5	Fable 5.1
Terminal-Bench-Science 0.1 (agentic scientific research)	24.7%	52.6%
Terminal-Bench 4.0 (agentic coding)	42.0%	55.8%
Humanity's Last Exam, with tools (multidisciplinary reasoning)	63.8%	65.0%

Read the spread, not the individual scores. Reasoning barely moved. Agentic coding moved a lot. Agentic science more than doubled. That distribution says the release was tuned for long-horizon tool use, the mode where a model works through a multi-step task in a terminal or lab pipeline rather than answering a single prompt. Anthropic's own examples point the same direction: protein binder design with affinities it claims are 10x higher than competing approaches, Venus elevation mapping refined from 10-20km resolution to 2-3km, and GPU kernel optimization delivering up to 2.5x inference speedups.

Who this release is actually for

Most people using Claude through the app will notice little. Fable is the top of the lineup, priced accordingly, and everyday chat, writing, and coding assistance were already well served by Sonnet 5 and Opus 5 at a fifth to half the token price. Check the AI Pricing Index for the current spread across the lineup.

The release matters for three groups:

1. Teams running long agentic pipelines. The 45% cost reduction on highly agentic work plus 75% cheaper cache reads changes the math on agents that loop for hours. If Fable 5 was too expensive to leave running, 5.1 is the number to re-check.
2. Research and biotech shops. A benchmark that doubles is unusual this late in a model generation, and Terminal-Bench-Science is the closest public proxy for automated lab and analysis work. The Life Sciences Verification Program, run in partnership with the US government, is expanding enrollment alongside the release.
3. Security teams. Anthropic claims 60% fewer false positives on cybersecurity tasks and stronger resistance to prompt injection, with a Cyber Verification Program gating the less-restricted Mythos variant to defensive professionals.

The safeguard split is the strategy

The two-model structure is worth a moment. Anthropic is shipping one set of weights behind two doors: a public one with standard safeguards, and a vetted one where cybersecurity and life sciences professionals get fewer refusals for legitimate dual-use work. Enterprise customers also get zero-data-retention deployment on customer-controlled cloud infrastructure under what Anthropic calls Enterprise Frontier Safeguards.

Fable 5.1 achieves 60% fewer false positives in cybersecurity tasks, with improved biology safeguards and enhanced robustness against prompt injection attacks. - Anthropic announcement

This is the same pattern Anthropic has been building toward all year: publish the capability, then meter access to the sharpest edges by verification rather than by blunting the model for everyone. It has already opened system-prompt shaping to developers and watermarked Claude's output; graduated access tiers are the third leg of that stool.

The practical takeaway

If you are choosing a daily driver for chat or writing, nothing changed this week; the ChatGPT vs Claude calculus is the same as it was in August. If you run agents against the Claude API, re-run your cost model with the new cache-read pricing before the end of the month, because a 75% drop on the line item that dominates agentic bills is the kind of change that quietly reorders the model leaderboard on cost per completed task. And if your work touches protein design, security tooling, or any pipeline that looks like Terminal-Bench-Science, the verification programs are now the gate between you and the strongest published numbers in that category.