Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on Sept. 1, 2026. VentureBeat reported the launch the same day. Anthropic says the two names are the same underlying model with different safeguard tiers: Fable 5.1 is generally available, and Mythos 5.1 is limited to trusted access programs for cybersecurity and life sciences work.
Platform docs list API model id claude-fable-5-1, a 1 million-token context window, a 128,000-token max output, adaptive thinking always on, default effort high, and a June 2026 knowledge cutoff. Input and output prices stay $10 and $50 per million tokens. Cache reads drop to $0.25 per million tokens, which Anthropic and VentureBeat both describe as a 75% cut from Fable 5's $1.00 cache-read rate. Five-minute and one-hour cache writes remain $12.50 and $20 per million tokens. Batch API input and output stay half price.
Company claim: Anthropic's launch table puts Fable 5.1 at 52.6% on Terminal-Bench-Science 0.1, 55.8% on Terminal-Bench 4.0 (60.9% for Mythos 5.1), 1,853 on GDPval-AA v2, 73.4% on CursorBench 3.2.0, and 31.4% on AutomationBench. Anthropic footnotes that production safeguards can zero some tasks and that its OSWorld 2.0 task release is not directly comparable to earlier published OSWorld numbers.
Independent: Artificial Analysis's live Claude Fable 5.1 model page, checked Sept. 14, 2026, shows an Intelligence Index score of 53 for Adaptive Reasoning, Max Effort, Default Fallback, about 67.1 output tokens per second, and about $7.63 cost per Index task, with roughly 190 million output tokens across the Index run. Artificial Analysis's Sept. 1 article, written against an earlier Index revision, reported a max-effort score of 66 and said Fable 5.1 still cost about 20% more per Index task than Fable 5 because it used about 1.7 times as many output tokens despite the cache-read cut. Treat the 66 and 53 figures as Index-version sensitive, not as a silent model change.
Contested science score: Anthropic reports 52.6% on Terminal-Bench-Science 0.1 in its own setup. The public Vals.ai Terminal-Bench Science board shows GPT-6 Astra at 65.71% and Claude Fable 5.1 at 34.29%. Do not collapse those into one ranking. Different harnesses, trials, and effort settings can move the same named benchmark.
Company science demos remain company claims unless an outside lab publishes its own methods write-up. Anthropic says Mythos 5.1 designed protein binders that external organizations validated, with about a 50% hit rate across 12 targets and affinities on three Adaptyv Bio competition targets about 10 times higher than the best prior competition designs. It also says Fable 5.1 built a higher-resolution Venus elevation map from Magellan radar data, released under a Creative Commons license, and that Mythos 5.1 sped up seven open-source biology models by up to 2.5 times with planned open-sourced kernels.
The Sept. 1 System Card is the primary safety artifact. Anthropic says Fable and Mythos share identical weights; Fable keeps dual-use biology and cyber blocks while Mythos relaxes those domain safeguards for vetted programs. The card judges chemical and biological risk at CB-1 and short of CB-2, keeps Fable on the same biology safeguards as Fable 5, and lowers the alignment catastrophic-harm assessment from very low to low after recent cyber-evaluation incidents. Fable 5.1 may discover software vulnerabilities in source code, while exploit generation and several dual-use cyber tasks stay redirected. Anthropic also pitches Enterprise Frontier Safeguards for customer-controlled monitoring data, with zero data retention for eligible customers until EFS rolls out.