Claude Fable 5.1: Smarter Coding, 75% Cheaper Cache Reads

Claude Fable 5.1: Smarter Coding, 75% Cheaper Cache Reads

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, about three months after the Fable 5 line shipped. The two names are one underlying model behind different safeguard layers: Fable 5.1 is generally available as claude-fable-5-1 on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry, while Mythos 5.1 stays restricted to vetted organizations in Project Glasswing.

Before you get excited about a brand-new architecture, though, read the pricing page first. The most interesting number here is not on the benchmark table — it's on the invoice.

The benchmark story: gains are real but lopsided

Anthropic's own numbers show the biggest jumps on long, multi-step agentic work rather than single-turn coding:

On Terminal-Bench-Science 0.1 — an agentic scientific-research benchmark — Fable 5.1 more than doubles Fable 5's score. On Terminal-Bench 4.0 it beats Opus 5, and the less-restricted Mythos 5.1 pushes that to 60.9%. Elsewhere the gains are far more modest: CursorBench 3.2.0 moves from 70.5% to 73.4%, and Humanity's Last Exam goes from 57.8% to 60.9% without tools.

Caveat before you screenshot these and win an argument: Anthropic reports a standard error of ±3.5–4.5 points per model on Terminal-Bench-Science, so treat the margin, not the ranking, with care. And these are vendor-measured numbers — every lab cherry-picks its leaderboard glow-ups. Still, that Terminal-Bench-Science jump matters if you care about agents that actually run research tasks start to finish, which is exactly the agentic scientific computing territory.

The invoice story: cache reads dropped 75%

Base pricing is unchanged — $10 per million input tokens, $50 per million output. The change is in cached input.

A cache read (where the model re-reads context it has already processed) drops from $1.00 to $0.25 per million tokens. That's 0.025x the base input price, versus 0.1x on every other Claude model. Anthropic estimates this cuts effective cost by roughly 25% for typical workloads and up to about 45% for context-heavy, tool-heavy agentic work.

Why does that matter so much for agents? Because persistent agents re-read the same giant prefix constantly — your codebase, system instructions, tool definitions, accumulated conversation history:

// Before (Fable 5): re-reading a cached prefix = $1.00 / MTok
// After (Fable 5.1): the same cached prefix = $0.25 / MTok
const resp = await anthropic.messages.create({
  model: "claude-fable-5-1", // GA since Sep 1, 2026
  max_tokens: 8192,
  system: [REPO_CONTEXT_SYSTEM], // expensive to encode, cheap to re-read
  messages: [{ role: "user", content: "implement the retry wrapper in db.ts" }],
  tool_choice: { type: "auto" }, // NOT { type: "any" } → returns 400
});

The economics shift is real: hourly cache writes stay at $20/MTok and 5-minute writes at $12.50, but every subsequent read is a quarter of the old price. For a Claude Code-style loop that re-reads the repo context on every step, that's the difference between "this is too expensive to leave running" and "this is the cheapest way to work."

Three breaking changes your agent stack will hit

Anthropic's "what's new" notes list three breaking changes, and the first one will bite immediately:

1. Forced tool use is gone. tool_choice set to any or tool returns a 400 on Fable 5.1. You must use auto with strict tool use, or structured outputs, instead.
2. Thinking blocks are model-bound. Fable 5.1 can read earlier models' thinking, but no earlier model can read its own. Router and fallback setups that switch down to an older model silently lose the reasoning trace.
3. Editing earlier turns invalidates thinking blocks. Injecting or deleting per-turn reminders, or rebuilding the system/tools arrays mid-conversation, now errors. The check applies to accounts created on or after August 31, 2026. The documented fixes are turn-scoped system messages and server-side context editing.

There are also documented regressions, which is refreshingly honest. Parallel tool calling is more variable — agent loops may fire one call per turn where Fable 5 batched several. The model narrates less, answers from memory more often at low effort, and prefers whole-file rewrites over targeted edits.

Safeguards: the same model, two unlock levels

Fable 5.1 runs under standard production safeguards. Cyber safeguards now permit vulnerability discovery but not exploit development, which Anthropic says cuts safeguard interventions in Claude Code by roughly 60% per session; biology safeguards fire 85% less on benign requests. Penetration testing, exploit generation, and binary-based vulnerability scanning still redirect to Opus.

Mythos 5.1 lifts some of those constraints for vetted defenders and life-sciences organizations. On the research side, Anthropic reports Mythos 5.1 designed protein binders at roughly 50% hit rate across 12 targets against a 10–15% norm, Fable 5.1 built a Venus elevation map at 2–3 km resolution, and custom GPU kernels sped up seven open-source genomics models by up to 2.5x. Claude Security, the product that scans codebases for vulnerabilities, is powered by Mythos 5.1 and available to all Claude Enterprise customers.

For enterprises there's also Enterprise Frontier Safeguards (EFS), a new architecture that lets organizations keep agent monitoring and evaluation data inside infrastructure they control rather than on Anthropic's side. And on the compliance front, all Fable 5.1 output carries a statistical text watermark required by the EU AI Act, with C2PA credentials on generated files; the watermark-detection API is in private preview.

Should you switch?

If you're running long-horizon agentic coding or research workloads, 5.1 looks like the biggest step up in the Claude lineup — and if you're spending a lot on cached context, the 75% cache-read cut pays for the migration by itself. If you're on simple single-turn completions, the gains are modest and the breaking changes are real friction. And whatever you do, run it as claude-fable-5-1 — tool_choice: { type: "any" } will not survive the upgrade. Compare it against Goose vs Claude Code territory before you commit your agent fleet, and keep an eye on how the benchmark arms race responds.

Sources

• Anthropic: Claude Fable 5.1 and Claude Mythos 5.1
• Anthropic Docs: Fable 5.1 model overview