Claude Fable 5.1 Doubles Science Scores, Cuts Costs

Claude Fable 5.1 Raises the Frontier and Lowers the Bill

Anthropic released Claude Fable 5.1 today alongside its restricted sibling Mythos 5.1, and the headline is an unusual combination: the new flagship posts its biggest gains on scientific research and agentic work while costing meaningfully less to run than the Fable 5 model it replaces. Anthropic bills the pair as its most advanced models for coding and knowledge work, and the benchmark deltas — some of them more than doubling the previous scores — back the claim up.

• Fable 5.1 scores 52.6% on Terminal-Bench-Science, more than double Fable 5's 24.7%, and jumps from 42.0% to 55.8% on agentic coding in Terminal-Bench 4.0
• API pricing holds at $10 per million input tokens and $50 per million output, but cache reads drop 75% to $0.25 per million — roughly 25% cheaper for typical workloads and up to 45% for agentic ones
• Available today on Claude.ai, the Claude API (model ID claude-fable-5-1), Amazon Web Services, Google Cloud, and Microsoft Azure
• Mythos 5.1 shares the identical underlying model but is limited to vetted organizations doing cybersecurity and life-sciences work through Anthropic's trusted access programs

What Does Claude Fable 5.1 Actually Improve?

The pattern across Anthropic's published numbers is consistent: the further a task strays from single-shot chat and into long-running, tool-using work, the bigger the gap over Fable 5. Business workflow automation on AutomationBench nearly doubles from 17.1% to 31.4%. Knowledge work measured by GDPval-AA climbs from 1,723 to 1,853. On Humanity's Last Exam, the multidisciplinary reasoning gauntlet, Fable 5.1 reaches 60.9% without tools and 65.0% with them.

The scientific research results are the ones I keep returning to. Anthropic says the model designed protein binders with roughly ten times the binding affinity of rival competition entries at about a 50% hit rate, reconstructed a high-resolution elevation map of Venus from 30-year-old NASA radar data, and accelerated seven deep learning models by up to 2.5x through GPU kernel optimization. One customer anecdote stands out on the engineering side: 9to5Mac reports that testing at the investment firm Millennium saw Fable 5.1 identify a rare system crash that had eluded both engineers and competing models for years.

How Much Cheaper Is Fable 5.1 to Run?

Sticker prices are unchanged, so the savings live in the cache. Reading previously processed context now costs a quarter of what it did, which matters enormously for agents that carry long histories through hundreds of tool calls. Anthropic's arithmetic puts typical workloads about 25% cheaper than Fable 5 and complex agentic tasks up to 45% cheaper, and Bloomberg's coverage leads with the same framing: cheaper and better at coding. Anthropic also notes that at low or medium effort settings, Fable 5.1 matches or beats Fable 5 outright — the efficient modes are no longer the compromise they used to be.

What Separates Fable 5.1 From Mythos 5.1?

The two models are the same system wearing different safeguards. Fable 5.1 ships broadly with standard protections, while Mythos 5.1 goes only to vetted organizations that need deeper access for defensive security or life-sciences research, gated through Anthropic's Cyber Verification and Life Sciences Verification programs. The safeguards themselves got smarter rather than stricter: Anthropic says cybersecurity classifiers now produce 60% fewer false positives and biology safeguards trigger 85% less often on benign queries — a meaningful quality-of-life win for the researchers we covered when Claude opened 10,000 free seats for scientists. It also lands the same week as Anthropic's published sandbox rules for AI evaluations, part of a broader pattern of the company documenting its safety plumbing in public.

The Takeaway

Frontier releases usually ask you to choose between capability and cost. Fable 5.1 is notable for refusing the trade: the largest benchmark jumps in the lineup arrive alongside a 75% cache-price cut aimed squarely at the agentic workloads where those gains matter most. The scientific research results hint at where this generation is headed — models as genuine research collaborators rather than autocomplete — and the rest of our artificial intelligence coverage will be tracking how quickly that promise turns into published results.

Sources: Anthropic — Introducing Claude Fable 5.1 and Claude Mythos 5.1 — September 1, 2026; Bloomberg — September 1, 2026; 9to5Mac — September 1, 2026.