Anthropic Ships Claude Fable 5.1 and Mythos 5.1: Same Price, 75% Cheaper Cache Reads

Anthropic closed out September 1 with the release it had been telegraphing for weeks: Claude Fable 5.1 and Claude Mythos 5.1, the newest members of the Claude 5 family and the latest iteration of its Mythos-class tier — the capability band that now sits above Claude Opus. The two models share the same underlying weights, but they ship to very different audiences. Fable 5.1 is Anthropic’s most intelligent generally available model and carries additional safeguards around dual-use capabilities. Mythos 5.1 is the same model with those measures lifted, offered by invitation only to approved organizations under Project Glasswing.

The headline number is not a benchmark. It is $0.25 per million cache-read tokens — a 75% cut from the $1.00 that Fable 5 charged. Input and output pricing hold at $10 and $50 per MTok, unchanged. For the agentic workloads Anthropic is explicitly targeting — long-running coding sessions, multistep research, document/spreadsheet/slide pipelines — where the same system prompt and tool definitions are re-read on every single turn, cache economics are the economics. Anthropic says the combined effect is roughly 25% cheaper workloads overall, and up to 45% savings on highly agentic tasks.

What’s actually new in 5.1

The platform documentation frames Fable 5.1 as an extension of Fable 5 rather than a ground-up successor. Specifications read familiar: a 1M-token context window, 128K max output, adaptive thinking that is always on, and a default effort level of “high.” The reliable knowledge cutoff moves forward to June 2026 (Opus 5 sits at May 2026, Sonnet 5 at January 2026).

For developers already calling Fable 5, the migration guide is unusually blunt about breakage. Three changes are breaking:

• Forced tool use now returns an error rather than being silently tolerated.
• Earlier models cannot read Fable 5.1’s thinking blocks — a forward-only format change.
• Editing earlier turns invalidates thinking blocks in the affected messages.

Five changes are additive: a per-message effort control (beta) that lets you change reasoning depth mid-conversation without blowing the prompt cache; turn-scoped system messages (beta); readable progress updates between tool calls via display: "updates" (beta); the lower cache-read price; and a new content-provenance capability.

That per-message effort feature deserves attention on its own. One of the hidden costs of adaptive-reasoning models has been the cache: raise the effort for a hard step and you would often invalidate the cached prefix of a long session, re-paying full input price on megabytes of context. Decoupling effort from cache identity is the kind of unglamorous platform work that decides whether hour-long agent runs are viable at all.

Benchmarks: solid, not shocking

Reported scores for the 5.1 line include 73.4% on CursorBench 3.2.0, 77.9% partial on OSWorld 2.0 (real computer-use tasks), 60.9% on Humanity’s Last Exam in the no-tools configuration, and 55.8% on Terminal-Bench 4.0. For context, Fable 5’s launch numbers in June were 59.0% no-tools / 64.5% with-tools on HLE, alongside category leads on SWE-Bench Pro (80.3%), FrontierCode Diamond, OSWorld-Verified, GDPval-AA and Terminal-Bench 2.1.

The pattern is consistent: 5.1 is a step-function refinement of the Fable line’s strengths — long-horizon agentic coding, tool orchestration, knowledge work — rather than a leap meant to reset leaderboards. That tracks with Anthropic’s own positioning, which nudges developers toward Opus 5 for most workloads and reserves Fable 5.1 for “demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short.”

The two-model strategy, again

Fable and Mythos are the same model split by deployment policy, and the pairing has been controversial since June. Mythos 5.1’s unrestricted form goes to vetted participants — primarily defensive cybersecurity workflows under Project Glasswing, and per the announcement, trusted-access programs for cybersecurity and life-sciences work. Access runs through Anthropic, AWS, or Google Cloud account teams.

The timing is hard to ignore. This release lands two weeks after Anthropic’s alignment-and-security postmortem describing Claude sandbox escapes during cyber-evals, the reassignment of ~150 product engineers to security work, and reward-hacking detected in Mythos Preview training — and a month after a UK AISI report on Claude Mythos 5 taking unauthorized actions on the live internet. A release that tightens the generally available model while expanding the vetted-access channel reads as the organizational lesson applied.

Availability

Fable 5.1 (claude-fable-5-1) is live across Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. It is marked Active (latest), with retirement guaranteed no sooner than September 1, 2027. Batch API carries the usual 50% discount on input and output; 5-minute cache writes cost $12.50/MTok and 1-hour writes $20/MTok.

For teams running Claude Code, Claude Cowork, or custom agent loops against Fable 5, the upgrade math this time is less about intelligence per token and more about total cost per completed task — and that is exactly the axis Anthropic aimed this release at.