Claude Fable 5.1: Cheaper Agent Runs, a Real Science-Benchmark Jump

What actually shipped

On September 1, Anthropic released Claude Fable 5.1 (generally available as claude-fable-5-1 on the Claude API, Bedrock, Google Cloud, and Microsoft Foundry) alongside Claude Mythos 5.1, which stays restricted to vetted organizations. Anthropic's own framing leaned heavily on "coding, knowledge work, and long-running problem-solving tasks" — the kind of multi-step, tool-using sessions that agentic dev workflows actually run, not single-shot chat replies.

The benchmark that matters more than the headline

The standout number is Terminal-Bench-Science 0.1, an agentic scientific-research benchmark first published August 27. Fable 5.1 scored 52.6%, more than double Fable 5's 24.7%, and ahead of Opus 5 (29.0%) and GPT-5.6 Sol (22.4%). On the broader Terminal-Bench 4.0 suite Fable 5.1 hit 55.8%. Those are real, sourced figures — but a fresh 0.1-version benchmark on a narrow domain is exactly the kind of number that's easy to overfit to before it's battle-tested across a full release cycle.

The part that actually changes how you build

For teams running agents in a loop — coding assistants, backtest orchestration, research pipelines — the more consequential change is pricing. Cache reads fall 75%, from $1.00 to $0.25 per million tokens. Anthropic estimates that translates to roughly 25% lower cost on typical usage and up to about 45% lower on agentic workloads, where the same context gets re-read across many tool calls. That's a direct lever on the unit economics of any always-on agent, not a benchmark bragging right.

Reasoning effort as a cost dial — and a warning label

Fable 5.1 exposes five reasoning-effort levels — low, medium, high, xhigh, max — with no option to disable reasoning entirely. Developer Simon Willison's now-annual "draw a pelican riding a bicycle" SVG test is a useful, if informal, illustration of the spread: the same prompt cost roughly 10 cents at low or medium effort, about 13 cents at high, $1.83 at xhigh, and $3.30 at max — for one SVG. Read the full test. Willison also noted the pelican test itself is showing weaker correlation to real model quality than it did in 2025, which is a fair reminder that any single benchmark — including Terminal-Bench-Science — is a signal, not a verdict.

One more change worth flagging

Anthropic also loosened Fable 5.1's cyber-safeguards to permit vulnerability *discovery* while still blocking exploit development, which is reported to have cut safety interventions inside Claude Code sessions by roughly 60%. For defensive security and code-review use cases, that means fewer false-positive refusals mid-session — a small but real quality-of-life change for anyone running the model against real codebases.

The takeaway for agentic builders

Don't route on the Terminal-Bench-Science headline alone — it's one narrow, brand-new benchmark. Do route reasoning effort deliberately: low or medium for routine codegen and boilerplate, and reserve xhigh/max for the handful of genuinely hard steps in a pipeline, since the cost curve is steep and non-linear. And if you're running long agentic sessions with repeated context (the common pattern in coding agents and multi-step trading-research pipelines), the cache-pricing cut is worth re-running your own cost benchmarks against — vendor percentages are estimates, not your invoice.