Anthropic Ships Claude Fable 5.1 and Mythos 5.1

Anthropic has introduced two models: Claude Fable 5.1 for all paying customers, and Claude Mythos 5.1 for vetted US organizations only. Both replace the 5 series at the top of the lineup. The stated focus is agentic coding, autonomous operation across multi-hour jobs, and scientific research. AWS published its companion post on September 1, 2026.

The numbers Anthropic reports

On Terminal-Bench 4.0, an agentic coding test, Anthropic reports 55.8 percent for Fable 5.1. For comparison it lists 42.0 percent for Fable 5 and 52.3 percent for Opus 5. Mythos 5.1 sits at 60.9 percent. On Terminal-Bench-Science 0.1 the company reports 52.6 percent for Fable 5.1 against 24.7 percent for Fable 5.

Other figures in the announcement: 1853 points on GDPval-AA v2 versus 1723 for Fable 5, OSWorld 2.0 at 77.9 percent partial and 41.7 percent strict, and Humanity's Last Exam at 60.9 percent without tools and 65.0 percent with tools. Every one of these numbers comes from the vendor's own announcement. No independent replication was available when this article was written.

What it costs

Anthropic lists $10 per million input tokens and $50 per million output tokens. Cache reads are priced at $0.25 per million tokens, which the company describes as a 75 percent reduction from Fable 5. Anthropic estimates savings of roughly 25 percent on typical workloads and roughly 45 percent on heavily agentic ones compared with Fable 5. These are vendor estimates, not measured results.

Where it runs

The API model ID is claude-fable-5-1. Anthropic names the Claude API, Claude Code, Claude Enterprise and the Claude Platform, plus Amazon Web Services, Google Cloud and Microsoft Azure. On AWS the model is addressed as global.anthropic.claude-fable-5-1. It is reachable through Amazon Bedrock, through AWS GovCloud (US), and through the Claude Platform in North America.

AWS classifies Fable 5.1 as a Covered Model: requests run in aws_review mode, and data may be retained up to 30 days with human safety review by Amazon personnel. Enterprise Frontier Safeguards customers can choose zero data retention through December 31, 2026. The AWS post does not state Bedrock rates; it points readers to the pricing page. Listed prerequisites are Python 3.10 or newer, Boto3, and the bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream permissions.

Mythos 5.1 stays behind a fence

Mythos 5.1 is available only to vetted US organizations admitted to one of two programs: Cyber Verification and Life Sciences Verification. As evidence of scientific value, Anthropic points to protein design work. The model produced binders with binding affinity 10 times higher than the best entries submitted to Adaptyv Bio competitions. Anthropic puts the hit rate at roughly 50 percent across 12 targets, against a typical 10 to 15 percent.

On safeguards, Anthropic claims 60 percent fewer false positives in the cybersecurity classifiers and 85 percent fewer refusals of benign biology requests than at the original Fable 5 launch. For the EU, outputs now carry a watermark. A detection API is offered to regulators, law enforcement, media, fact-checkers, researchers and educators. Enterprise Frontier Safeguards are set to roll out in phases starting in fall 2026.

Reading the release

The jumps on the agentic benchmarks are large, but all of them come from the vendor's announcement. Whether 55.8 percent on Terminal-Bench 4.0 survives contact with production workloads is untested. AWS frames the gain more cautiously: better judgment on ambiguous tasks and fewer confident wrong answers. That is the more checkable claim, and it can only be measured in use.