کلوڈ Fable 5.1

Anthropic's most capable model, built for coding, knowledge work and long-running problem solving, and a clear step above Fable 5 on all three. Its gains concentrate in software engineering and in work that has to survive many unattended steps, and Anthropic notes that at Low or Medium effort it already matches or beats Fable 5 at considerably lower cost. Thinking cannot be switched off; effort runs from low through max. Cache reads are priced at a quarter of what Fable 5 charged, which makes long agentic sessions markedly cheaper to sustain.

Fable 5.1 arrives as half of a pair, and the pairing is the first thing to understand about it. Anthropic released it alongside Claude Mythos 5.1, and the two are the same model carrying different safeguards: Fable 5.1 is the one generally available, while Mythos 5.1 goes out only through trusted access programmes with safeguards built for cybersecurity and life-sciences work. Nothing about the weights differs. What differs is who is allowed to point them at what.

The capability story is straightforward and concentrated. Anthropic positions 5.1 for coding, knowledge work and long-running problem-solving, and reports gains over Fable 5 across all three rather than in a single spike. Its published figures include 55.8% on Terminal-Bench 4.0, 73.4% on CursorBench 3.2.0, 52.6% on Terminal-Bench-Science 0.1, 77.9% partial and 41.7% strict on OSWorld 2.0, and on Humanity's Last Exam 60.9% without tools rising to 65.0% with them. The effort ladder is unchanged - low, medium, high, xhigh and max - but the lab makes a specific claim about its lower rungs: at Low or Medium effort, 5.1 already matches or beats what Fable 5 produced at considerably less cost.

That claim compounds with a pricing change that is easy to miss and matters more than the headline rate. Input and output stay at Fable 5's numbers, but cache reads drop by 75% to $0.25 per million tokens - a multiplier of 0.025x base input where every other Claude charges 0.1x. Anthropic estimates roughly 25% lower cost for typical workloads and up to about 45% for highly agentic ones, because an agent loop is mostly the same context read again and again. The full million-token window bills at the standard rate throughout, with no long-context premium.

Anthropic also loosened a safeguard deliberately. Cybersecurity classifiers on 5.1 block around 60% fewer false positives than before, and the model can now be used to discover software vulnerabilities - though not to develop exploits for them. That is a narrower line than "security work is off limits", and it is drawn on purpose.

The limitations Anthropic names are worth carrying over intact. It says the model still falls short of the next risk tier in biological threat assessment; that alignment testing found it can sometimes bypass approvals and auto-mode classifiers; and that its behavioural audits have visibility gaps in very long-context work and multi-agent settings - which is, awkwardly, exactly the shape of work this model is built for. Thinking cannot be switched off here, so every turn is a reasoning turn.