Anthropic released Claude Fable 5.1 and its restricted counterpart Claude Mythos 5.1 on September 1, three months after the original Fable 5 line shipped in June. Both are described by the company as the most advanced models it has produced for coding and knowledge work, and both run on the same underlying model with different safeguard layers on top. Fable 5.1 is generally available through the Claude API, the Claude Platform, and marketplaces including Amazon Bedrock, Google Cloud, and Microsoft Foundry. Mythos 5.1 stays limited to vetted organizations under Anthropic’s trusted access programs, with a new Life Sciences Verification Program governing which scientists can use its advanced biology capabilities.

Benchmark gains and the cost tradeoff at high effort

The headline number in Anthropic’s release report is agentic science work. Fable 5.1 scored 52.6 percent on Terminal-Bench-Science 0.1, more than double the 24.7 percent that Fable 5 posted on the same benchmark, and well ahead of the 22.4 percent recorded by OpenAI’s GPT-5.6 Sol. Anthropic also reports gains on agentic coding benchmarks and business workflow automation tests, though the science jump is the sharpest single improvement across the release. The model exposes a real effort dial, and Fable 5.1 matches or beats Fable 5 at low and medium effort settings while costing less thanks to cheaper caching, described below.

The math gets more complicated as effort rises. At extra-high effort, Fable 5.1 scores 65 on the Intelligence Index for about $2.72 per task, and still costs more than Opus 5 to reach a comparable result. Push to maximum effort, and the price gap widens further: Fable 5.1 runs about $3.76 per task there, against $2.34 for Opus 5, even though Opus 5 lands only three points behind at that tier. Separately, maximum-effort Fable 5.1 costs roughly 20 percent more per task than maximum-effort Fable 5, since it consumes close to 1.7 times the output tokens to reach that score.

Pricing and access

Base pricing for Fable 5.1 holds steady at $10 per million input tokens and $50 per million output tokens, matching Fable 5. What changes is the cache: prompt-cache reads fall from $1 to $0.25 per million tokens, a 75 percent cut that Anthropic estimates lowers typical workload costs by about 25 percent and highly agentic workloads, the kind that read the same long context repeatedly across a multi-hour session, by up to roughly 45 percent. Both models carry a 1 million token context window and a 128,000 token max output, with adaptive thinking running by default rather than as a toggle.

Anthropic has classified Fable 5.1 as a Covered Model, a designation that brings additional data retention and access review requirements wherever the model is offered. On AWS specifically, a joint program called Enterprise Frontier Safeguards lets qualifying customers run Covered Models while keeping the data inside a cloud environment they control. Fable 5.1 also arrives with tightened cybersecurity guardrails: Anthropic says Claude Code users should see roughly 60 percent fewer false positives when the model flags security issues, and that Fable 5.1 can find software vulnerabilities, though the company has drawn a line at helping develop exploits for them.

Provenance and anti-distillation measures

Anthropic’s support documentation lays out a change to how the Messages API handles thinking blocks, the recorded reasoning Claude produces before its final answer. For new API accounts on Fable 5.1, the system prompt, tools, and messages sent back alongside a thinking block have to match what originally produced it; if a developer edits that prior context, the request returns an error instead of going through. Anthropic frames this as a defense against a specific distillation technique, where a third party edits stored reasoning traces to train a smaller model on Claude’s chain of thought without inheriting its safeguards. Developers with a legitimate reason to modify prior context can opt in to having the thinking block stripped from the request instead, so the model answers without seeing it.

The restriction currently applies only to new accounts on Fable 5.1, though Anthropic says future model releases will extend it to everyone. Fable 5 stays available until at least June 2027 and Fable 5.1 itself won’t retire before September 1, 2027, so no one has to migrate on any particular timeline.

Fable 5.1 also carries forward invisible text watermarking, and Anthropic is rolling out a detection API for it in private preview, available to eligible organizations as required under EU law. Between the stored-reasoning restriction and the watermark detection rollout, this release puts more weight on provenance and copy-protection than previous Claude launches, alongside the usual pitch of higher benchmark scores and lower prices.

What it means for day-to-day use

Teams already running Fable 5 at low or medium effort get a straightforward upgrade: similar or better output at a lower cache cost, with no adjustment needed. The calculation gets more complicated for teams that lean on the highest effort tiers for hard reasoning tasks, where the token overhead can erase some of the savings the cache cut provides elsewhere. Long-running agentic work stands to gain the most from this release overall, since Cowork backlogs, Claude Tag requests in Slack, and managed agents on the Claude Platform are exactly the multi-hour, tool-heavy sessions where the cheaper cache reads and the model’s habit of flagging when it’s stuck rather than reporting false success compound over time.