Anthropic Ships Claude Fable 5.1, Cuts Agent Costs Up to 45%
The new model beats Fable 5 on every published benchmark and comes with cheaper caching, while a restricted sibling model targets cybersecurity and life-sciences work.
Anthropic released Claude Fable 5.1 on September 1, 2026, three months after Fable 5, along with a more tightly access-gated sibling called Claude Mythos 5.1. Anthropic is calling them the company's most capable models yet for coding and knowledge work, and the benchmark numbers back that framing up more than most release-day claims do.
What actually improved
The headline number is Terminal-Bench-Science 0.1, a test of agentic scientific research tasks. Fable 5.1 scored 52.6% there, more than double Fable 5's 24.7% and well ahead of Opus 5's 29.0%. On Terminal-Bench 4.0, a broader coding and terminal-use benchmark, Fable 5.1 hit 55.8%. On Humanity's Last Exam, a notoriously hard general-knowledge test, it scored 60.9% without tool use and 65.0% with tools enabled.
Those aren't small gains dressed up with a version bump. Anthropic says Fable 5.1 finishes ahead of Opus 5 on every category it published results for, which is a real claim worth checking against independent evals as they come in over the next few weeks.
Cheaper, not just faster
Pricing on paper hasn't moved. Input tokens are still $10 per million and output tokens $50 per million, the same list price as Fable 5. What changed is cache-read pricing, cut 75% to $0.25 per million tokens. For a typical workload that's roughly a 25% cost reduction. For heavily agentic workloads, the kind that lean hard on cached context across long multi-step tasks, Anthropic says the savings run up to 45%. If your bill is dominated by an agent re-reading the same codebase or document set turn after turn, this is the part of the release that actually shows up on an invoice.
Two models, two audiences
Fable 5.1 is generally available today through the Claude API as claude-fable-5-1, plus Claude.ai, Claude Code, Claude Enterprise, and the usual cloud triumvirate of AWS, Google Cloud, and Microsoft Azure. Mythos 5.1 is the same underlying model with different safeguards, and it's not generally available. Anthropic is routing it only to vetted cybersecurity professionals through a Cyber Verification Program and to life-sciences researchers through a separate verification program, specifically because those fields need a model willing to discuss exploit mechanics or pathogen biology in ways Anthropic doesn't want available to anyone who signs up for an API key.
The safeguard tuning on the generally-available Fable 5.1 also got more precise, not just looser. Anthropic says biology-related safety triggers fire 85% less often on benign requests, and cybersecurity false positives dropped around 60%, while still restricting offensive exploit generation. That's the harder problem in AI safety engineering: not being cautious, but being cautious about the right things without blocking a security researcher trying to write a proof-of-concept for a bug they already found.
The verdict
A doubling on a science-agent benchmark and a real double-digit cost cut in the same release is a genuinely strong showing, not just an incremental point release despite the ".1" naming. If you're already building on Fable 5, swapping to 5.1 looks like a straightforward upgrade with no price increase and better numbers across the board. The bigger question is whether Terminal-Bench-Science and CursorBench scores translate into fewer failed agent runs in production, and that's something only a few more months of real usage will answer.