Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1 — this is one model with two levels of protection: Fable 5.1 is open to everyone on all platforms, while Mythos 5.1 is only available through verified access programs for cybersecurity and life sciences. Base rates remain the same, but cache reads have become four times cheaper, so bills for agentic scenarios may drop by almost half. According to vendor measurements, Fable 5.1 outperforms both Fable 5 and Opus 5, while operating approximately twice as fast as Opus 5.

image
image

What happened

Anthropic split the release into two versions of one model. Claude Fable 5.1 is available to everyone on all platforms, its API identifier is claude-fable-5-1; Claude Mythos 5.1 differs in relaxed safeguards and is only available to participants in verified access programs in cybersecurity and life sciences. Input and output prices have not changed — 10 and 50 dollars per million tokens, — and cache reads have become cheaper from one dollar to 0.25 dollars per million tokens. According to Anthropic's measurements over four weeks in August 2026, this yields about 25% savings on standard tasks and up to 45% on agentic tasks, where cache reads account for up to 65% of the bill, because the agent re-reads the same context at each step. In benchmarks, Fable 5.1 scores 55.8% on Terminal-Bench 4.0 (Mythos 5.1 — 60.9%), 52.6% on Terminal-Bench-Science 0.1 versus 24.7% for Fable 5, 65.0% on Humanity's Last Exam with tools, and 73.4% on CursorBench 3.2.0. On the day of the release, Cognition announced the migration of Opus 5 traffic in Devin to Fable 5.1, and EFS was also announced, which moves data storage to the client's infrastructure.

Context

Caching in large language model APIs allows re-reading already processed context significantly cheaper than processing it anew, and this is what the economics of agentic systems is based on: an agent re-reads the same context at each step of the pipeline, so in agentic scenarios cache reads reach 65% of the final bill versus approximately 40% on standard tasks. The price of such reads before this release was significantly higher than the base processing cost, and its reduction affects the final bill more than any changes to base rates, which were not touched at all this time. Splitting one model into a publicly available Fable and a gated Mythos with relaxed safeguards — is a continuation of Anthropic's policy of controlled access to strong models in sensitive areas: a more free version is only available to verified organizations in cybersecurity and biology, while the general public receives a version with stricter protections.

Why this matters for the industry

For builders of agentic systems, this is a rare case where unit economics improve without any changes to the product: the vendor did not touch input and output prices, but cut the most expensive part of the agentic bill by 75%. Long multi-step products — coding agents like Devin, enterprise pipelines with reusable context — become significantly cheaper, and scenarios that previously did not work out in unit economics may become profitable. Cognition has already migrated Devin to Fable 5.1 on the day of the release, showing how major consumers evaluate this migration. The public delta of 55.8% versus 60.9% between Fable 5.1 and Mythos 5.1 on Terminal-Bench 4.0 — is a rare quantitative assessment of how much capability the tightening of safeguards costs, and a useful signal about the trade-off between security and utility. If the four-fold reduction in cache costs takes hold, competitors will likely begin to repeat this step, which puts pressure on the economics of all agentic systems and shifts prompt design toward cache-friendly architectures with stable prefixes and persistent project context.

Why this matters for users

If you work with the Claude API, Claude Code, or agentic pipelines, migration reduces to changing the model identifier to claude-fable-5-1: you do not need to change the tariff plan, new prices are already in effect. The real gain depends on what share of your bill is occupied by cache reads, so the first practical step is to take this share from billing, and the second is to run your own evals on Fable 5.1 against the current model before moving production. According to vendor measurements, the model is approximately twice as fast as Opus 5; effort by default is set to High in Claude Code and to Medium on claude.ai, so behavior can be additionally adjusted to the task. The third step is to move stable context to a cacheable prefix: the more reusable context in your pipeline, the stronger the effect of the new cache read price on the final bill.

What is still unknown / limitations

All figures on the model's capabilities are self-measurements by Anthropic without disclosed methodology, and there is no independent confirmation of the increase yet. The jump on Terminal-Bench-Science 0.1 from 24.7% to 52.6% within a minor update from 5 to 5.1 is abnormally large for a point release and is the first candidate for independent reproduction: either the increase is real, or the result is highly sensitive to the format and prompts of the benchmark. The difference between Fable 5.1 and Mythos 5.1 on Terminal-Bench 4.0 — is also a vendor measurement. The scientific novelty of the minor release 5.1 is not shown in the announcement. The expected savings are checked by each client's own billing and depend on the structure of its expenses, so there is no universal figure for all loads.

Sources

Author

Look at AI, editorial team