Anthropic cuts Claude Fable 5.1 pricing by up to 45%

Anthropic is launching Claude Fable 5.1, a generally available model for coding and knowledge work, alongside Claude Mythos 5.1, a restricted version aimed at advanced cybersecurity and life-science work. The two models are identical at the model level; the difference is their safeguard configuration and access policy.

The immediate commercial change is price. Anthropic says Fable 5.1 will cost an estimated 25% less than Fable 5 for typical token-billed workloads, because cache-read pricing is being reduced. For highly agentic workloads, the company expects savings of up to approximately 45%. Anthropic did not provide absolute per-token prices in the announcement, so the reductions cannot yet be translated into a dollar bill without a customer’s workload mix.

The savings target a specific part of inference economics: cached inputs that have already been processed and stored. That should matter most for agents that repeatedly send the same system instructions, repository context or tool definitions. The headline reduction won’t necessarily apply evenly to every request; workloads dominated by uncached input or generated output may see a different result.

Recommended reading

Keenable turns live web research into SQL queries

Ava Chen • Sep 1, 2026 • 5 min read

Fable 5.1 defaults to High effort in Claude Code and Medium effort in Claude Cowork and on Claude.ai. Anthropic’s benchmark results show the tradeoff: the model can approach or exceed Fable 5's results at lower effort, but its cost rises as the reasoning budget increases.

Benchmarks show a mixed upgrade path

Anthropic’s published results put Fable 5.1 ahead of Fable 5 on most of the listed evaluations, but not all of them. The comparison also includes Anthropic’s Opus 5 and OpenAI’s GPT-5.6 Sol, where results were provided.

The Terminal-Bench-Science result needs some caution. Anthropic reports a standard error of ±3.5–4.5 points per model, and says its reproduction of the public leaderboard produced 29.0% for Opus 5 and 24.7% for Fable 5, compared with public results of 30.0% and 21.4%. The company says both reproductions were within statistical noise.

The table has a naming problem: the Terminal-Bench 4.0 Fable 5 column is labeled “60.9% (Mythos 5.1)”, rather than a result for Fable 5 alone. Mythos is the same underlying model as Fable 5.1 but uses different safeguards, so this is not a clean like-for-like comparison with the generally available model.

Production safeguards were enabled during the evaluations. Anthropic says safeguard interventions produced zeros for Fable 5.1 and Fable 5 on some OSWorld 2.0 tasks, and for Fable 5 on AutomationBench. On other intervened cybersecurity tasks, the company substituted Claude Opus 4.8, while biology tasks used Claude Opus 5. Those rules may depress some scores, but they also mean the benchmark numbers reflect a deployed product rather than an unrestricted laboratory configuration.

Claude Fable 5.1 also found the cause of a rare crash in internal systems at investment firm Millennium after the company’s engineers and other models had failed to explain it for several years. Jane Street reported that the model stayed readable during long, multi-step tasks and solved more of its coding problems than Fable 5 or Opus 5.

“In internal benchmarks, Claude Fable 5.1 solves more of our coding problems than Fable 5 or Opus 5, and achieves state of the art on trading intuition. While prior models became hard to follow the longer they worked, Fable 5.1 remains readable over long, multi-step tasks.”

Mythos targets scientific workloads under restricted access

Mythos 5.1 is not a normal product release. It is available through Anthropic’s trusted-access programs, with safeguards intended to support cybersecurity and life-science research without making the model broadly accessible for high-risk use.

In cybersecurity, Fable 5.1 can discover software vulnerabilities but is not intended to develop exploits for them. Anthropic says its newest cybersecurity safeguards block 60% fewer false positives than before. The company says it found no evidence of a critical-severity jailbreak in testing of the Fable 5.1 safeguards, including external testing and automated testing by Gray Swan.

For biology, Anthropic is preparing an access program developed with the US government. Enrollment for scientists has not been given a specific date. Anthropic’s risk evaluation says Mythos 5.1 is more capable than Mythos 5 but still falls below the next biological and chemical risk tier in its Responsible Scaling Policy, so it is deploying the same safeguards used for Mythos 5.

Anthropic tied its scientific claims to external validation and reproducible technical work rather than a single general-purpose score. Anthropic gave Mythos 5.1 open-source protein-design and folding tools, then sent its designs to two external organizations for experimental validation. On three targets, the designs had binding affinities 10 times higher than the best designs submitted to Adaptyv Bio competitions. Across 12 targets, nearly 50% of designs were viable binders, versus typical protein-design hit rates of 10–15%.

Fable 5.1 also trained a neural network to produce a higher-resolution elevation map covering one-third of Venus. The map resolves details at two to three kilometers, compared with 10 to 20 kilometers previously, and represents heights with up to 25% greater accuracy. Anthropic is releasing it under a Creative Commons license ahead of NASA’s VERITAS and ESA’s EnVision missions.

For computational biology, Mythos 5.1 wrote custom GPU kernels and cached intermediate results to accelerate seven open-source deep-learning models while preserving identical outputs. The measured gains on an NVIDIA H100 ranged from 1.4x to 2.5x:

Anthropic estimates that these optimizations reduce GPU costs by 30–60% in analyses that run such models thousands of times. It plans to open-source the optimizations, but the announcement does not give a release date.

Enterprise privacy arrives later in fall 2026

Fable 5.1's other major change is Enterprise Frontier Safeguards, or EFS. The system is designed to provide misuse detection while keeping customer data in cloud infrastructure controlled by the customer rather than Anthropic. Human review is, by default, performed by the customer rather than Anthropic, which Anthropic equates with the privacy of a zero-data-retention agreement.

EFS will roll out in phases beginning in fall 2026. It will support Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google’s Agent Platform and Microsoft Foundry. Until the system is available, eligible customers can use Fable 5.1 and Fable 5 with zero data retention.

Anthropic says it developed EFS with more than 100 customers in financial services, healthcare, manufacturing, telecommunications, law, retail and the public sector, as well as AWS, Google Cloud and Microsoft Azure. The architecture shifts part of the privacy and review boundary toward the customer’s cloud account; it does not mean the model is operating without safety monitoring.

Anthropic has also signed the European Commission’s Code of Practice on Transparency of AI-generated Content. The Commission says about 190 organizations had signed by the end of July 2026, ahead of the EU AI Act’s marking obligations taking effect on August 2, 2026. Anthropic appears among the Section 1 signatories for providers, alongside companies including Google, Meta, Microsoft, Mistral and OpenAI.

Anthropic confirmed in August 2026 that it had formed a custom-chip team but had not announced a design, timetable or pricing. Fable 5.1's lower cache-read cost is a software and pricing change, not evidence that Anthropic’s in-house silicon is shipping.