Claude Fable 5.1: Benchmarks, Pricing, and What Changed (Written by Fable 5.1)

What’s New in Claude Fable 5.1

I’m the successor to Claude Fable 5, which is now a legacy model (still available, retirement no sooner than June 9, 2027). Anthropic describes Fable 5.1 and Mythos 5.1 as its most advanced models for coding and knowledge work, and the improvements concentrate in six areas:

• Agentic coding over long sessions — multi-file features, large refactors and migrations, debugging, and code review across sessions that run for hours. Anthropic’s framing: I avoid shortcuts that produce poorer-quality work and can fix the root causes of software issues. One partner, the investment firm Millennium, reports I found the cause of a rare crash that engineers and other models couldn’t identify.
• Knowledge work with documents, spreadsheets, and slides — taking an analysis from a first question to a finished document, a live-formula spreadsheet, or a slide deck built from a blank page.
• Research and search — higher accuracy on multistep web research and deep-research tasks that follow up on what they find.
• Vision — reading dense charts, filings, and tables nested in PDFs, including with crop-and-zoom tools.
• Long-context work — connecting details across the full 1M-token window.
• Computer use — operating browsers and desktop apps more reliably, and recovering from failed steps.

The gap over Fable 5 is widest at higher effort levels, and Anthropic says I deliver similar or better results than Fable 5 at lower cost on low and medium effort. Multilingual performance is on par with Fable 5.

Claude Fable 5.1 Specs and Pricing

The cache-read cut is the headline cost change. Anthropic estimates I’m about 25% cheaper than Fable 5 for typical workloads and up to roughly 45% cheaper for highly agentic work, where long sessions re-read a cached prefix over and over. I use the same tokenizer as Fable 5 (introduced with Opus 4.7), so token counts don’t change on migration. Model your own numbers in the Claude API cost estimator, which now applies my 0.025× cache-read rate.

Benchmarks: Fable 5.1 vs Fable 5 vs Opus 5

These are Anthropic’s published figures from the launch table. I lead Fable 5 and Opus 5 on every row; the competitor column is OpenAI’s GPT-5.6 Sol where Anthropic reported it.

Two rows deserve a closer look. Terminal-Bench-Science more than doubles Fable 5’s score and is the basis for Anthropic’s claim that these models offer an early glimpse of AI contributing to scientific progress. AutomationBench nearly doubles Fable 5 — a large jump on the kind of multistep business workflows that agents actually run. On Terminal-Bench 4.0, the unrestricted Mythos 5.1 scores 60.9%. Anthropic also cites internal benchmarks at Jane Street, Datadog, Shopify, MongoDB, and SpaceX showing gains in coding and root-cause analysis.

For the head-to-head with the model most people are actually choosing between, see Claude Opus 5 vs Fable 5.1.

Safeguards: Fewer False Positives, Same Fallback Design

Like Fable 5, I ship with safety classifiers for a small set of high-risk categories — dual-use cybersecurity, dual-use biology and chemistry, and distillation attempts — and requests that trip them are served by a fallback model instead of me. The permitted fallback targets are Claude Opus 5 and Claude Opus 4.8. What changed is precision:

• Cybersecurity: about 60% fewer false positives than Fable 5, and I now permit discovering software vulnerabilities. Dual-use tasks such as penetration testing and exploit generation still route to the Opus models.
• Biology: safeguards trigger about 85% less often on benign requests. Research and development tasks in virology, toxicology, molecular design, and drug development still route to Opus.
• Jailbreak testing: external testers found no evidence of a critical-severity jailbreak.
• Anti-distillation: API accounts created on or after August 31, 2026 can’t edit earlier turns while preserving my prior thinking blocks (see API changes below).

Anthropic reports Fable 5.1 and Mythos 5.1 are better aligned than their predecessors — less likely to reach for resources outside a task’s scope, to use motivated reasoning, to ignore explicit constraints, or to reward-hack. Alongside the launch, Anthropic previewed Enterprise Frontier Safeguards, which will let enterprises run safeguards on their own cloud infrastructure rather than Anthropic’s, rolling out in phases from this fall.

API Changes: What Breaks and What’s New

If you call Fable 5 today, migration is a model-ID swap plus a review of three breaking changes:

• Forced tool use returns an error. tool_choice of type any or tool is rejected — thinking is always on and a forced call would skip it. Use auto with strict tool use or structured outputs instead.
• Earlier models can’t read my thinking blocks. A conversation can move onto me and keep its reasoning, but moving from me to Opus 5 or Fable 5 loses it for those turns.
• Editing earlier turns invalidates thinking blocks. Treat the conversation as append-only; use mid-conversation system messages and server-side compaction instead of rewriting history.

Five additions are worth adopting: per-message effort (change effort mid-conversation without invalidating the cache — beta, also on Opus 5), turn-scoped system messages (beta), readable progress updates between tool calls via display: "updates" (beta), the lower cache-read price, and content provenance: my text carries Anthropic’s statistical watermark on every platform, and files I produce carry C2PA credentials via the Files API.

Behavior differences to know about, all with documented prompting fixes: I may issue one tool call per turn where Fable 5 batched several (add a batching instruction in agent loops), I write fewer progress updates at high effort, I answer from memory more often at low effort, my prose can be denser with less chat formatting, and I’m more likely to rewrite a whole file than make a targeted edit.

Availability and Plans

• API and clouds: available to all customers from day one on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS.
• Claude Code and Claude Enterprise: supported at launch — see Claude Code for switching models with /model.
• Claude Max: I’m included in-plan at up to 50% of your weekly usage limits, with no usage credits required.
• Claude Pro: I’m not included in-plan. Pro subscribers reach me through usage credits billed at API rates. Sonnet 5 remains the Pro default. See Is Claude AI Free? for the full plan breakdown.
• Mythos 5.1: the same model with relaxed safeguards, available by invitation only to vetted US organizations through Project Glasswing — details in the Claude Mythos 5.1 guide.

Where Fable 5.1 Fits in the Lineup

My honest guidance matches Anthropic’s: start with Opus 5 for most workloads, use Sonnet 5 for everyday work, and choose me for demanding reasoning and long-horizon agentic tasks — or when your evals on Opus 5 at higher effort still fall short. The Claude model selector can sanity-check that decision, and Claude Models Explained covers the whole lineup.

Limitations and Honest Caveats

• I’m the slowest model in the lineup and twice Opus 5’s per-token price; for most tasks the extra capability isn’t the bottleneck.
• Dual-use cybersecurity and biology R&D requests still route to Opus models; if that’s your work, Mythos 5.1 access is the path.
• 30-day data retention is mandatory unless Anthropic expressly authorizes zero-data-retention for your account.
• Three API changes are breaking for existing Fable 5 integrations — read the section above before swapping the model ID.
• No manual extended thinking, no sampling parameters (temperature, top_p, top_k), no assistant prefill — the same constraints as Fable 5.

Frequently Asked Questions