Anthropic Releases Claude Fable 5.1 and Mythos 5.1

Anthropic has made Claude Fable 5.1 generally available and is limiting its identically trained sibling, Mythos 5.1, to two vetted access programs. The company describes them as one model shipped with two different levels of safeguards. Fable 5.1 is reachable through the Claude API as claude-fable-5-1 and through Amazon Web Services, Google Cloud and Microsoft Azure.

Cheaper, and cheapest for agents

List pricing is $10 per million input tokens and $50 per million output tokens. Cache reads land at $0.25 per million tokens, a 75% cut. Anthropic estimates savings of around 25% on typical workloads and up to approximately 45% on agentic work. Those figures come from the vendor; no independent cost measurements are available yet.

Reading the benchmark table

The widest gap Anthropic reports sits on a command-line benchmark aimed at scientific work. The remaining scores cluster within a few percentage points of the previous generation.

• Terminal-Bench-Science 0.1: 52.6% versus 24.7% (Fable 5) and 29.0% (Opus 5)
• Terminal-Bench 4.0: 55.8% versus 42.0% and 52.3%
• OSWorld 2.0 (partial run): 77.9% versus 72.9% and 75.4%
• Humanity's Last Exam with tools: 65.0% versus 63.8% and 63.6%
• CursorBench 3.2.0: 73.4% versus 70.5% and 70.0%

The science claims

Anthropic offers three examples from the natural sciences. In protein design the model reportedly reaches a hit rate of close to 50% across 12 targets. A reprocessing of Venus elevation data is said to expose features down to 2 to 3 kilometers, against 10 to 20 kilometers before. In computational biology the company puts the speedup on certain deep learning models at up to 2.5 times.

Two models, two doors

Mythos 5.1 stays behind two verification programs: one for defensive security work, and one for life sciences run with the US government and currently restricted to the United States. Anthropic also reports fewer misfires from its filters, with 60% fewer false positives in the cybersecurity layer and 85% fewer refusals on benign elementary biology questions. Enterprise customers can additionally use Enterprise Frontier Safeguards, which keep data in cloud infrastructure the customer controls.

What the AWS route looks like

On AWS the model carries the ID global.anthropic.claude-fable-5-1 and has been available since September 1, 2026 on Amazon Bedrock and Claude Platform on AWS, including GovCloud endpoints. AWS classifies it as a Covered Model, which adds rules for data retention, safety review and access. By default AWS keeps prompts and outputs for up to 30 days with human safety review, and the retention mode has to be set through the data retention API before the first call. Eligible customers can run with zero data retention through December 31, 2026.

What stays unverified

The AWS post points to the Bedrock pricing page rather than quoting rates, so the two sources read here do not establish whether AWS pricing matches Anthropic's list prices. Neither source documents how the benchmark runs were configured. Anthropic mentions invisible watermarking framed around the EU AI Act but does not describe a way for third parties to check it.