Overview
Claude Fable 5.2 is Anthropic's November 9, 2026 upgrade for difficult reasoning, coding and long agent runs. It lowers standard API token rates by 20% from Fable 5.1 and improves several Anthropic evaluations. A model-ID change is enough for a simple request. An application that stores or edits multi-turn conversations, uses an older computer-use tool, or pairs two model versions as executor and advisor needs a migration test.
What's New
Improves results on difficult agent tasks
Anthropic reports 49.1% on FrontierCode 2.0 at Xhigh effort, versus 32.4% for Fable 5.1 at the same effort. AutomationBench rose from 31.4% to 51.4%. These are Anthropic's reported evaluations. Use them to pick workloads to retest, and measure success and cost in your own harness.
Reduces standard API rates and adds Fast Mode
Standard input, output and cache-read rates are 20% lower than Fable 5.1. Anthropic reports up to 30% faster output generation than the previous model and up to 2.5× token generation speed in optional Fast Mode. Token generation speed measures only one part of a tool-heavy job.
Reduces safeguard interventions in early testing
Anthropic observed about 55% fewer cyber and biology blocks combined across Claude's web app, Code and Cowork. High-risk activities can still trigger safeguards and fallbacks. If your workflow requires vetted access to those capabilities, check eligibility for Mythos 5.2 through Anthropic’s verification programs.
Compared With Previous Version
| API consideration | Fable 5.1 | Fable 5.2 |
|---|---|---|
| Standard input / output per million tokens | $10 / $50 | $8 / $40 |
| Cache reads per million tokens | $0.25 | $0.20 |
| Context / maximum output | 1M / 128K tokens | 1M / 128K tokens |
| Conversation-prefix check | Enforced by default for newer accounts | Applies to every account |
| Rate limit | Shared with Fable 5 | Separate limit for Fable 5.2 |
On the Claude API, Fast Mode is $16 input and $80 output per million tokens. Compare elapsed time and spend per completed task before applying it across a queue. Anthropic lists eligible platforms and platform-specific configuration in its migration guide.
Migration Guide
- Check workspace and capacity requirements before switching. On the Claude API, Fable 5.2 requires 30-day data retention unless Anthropic expressly authorizes zero data retention. It has its own rate limit and does not support Priority Tier.
- Set the Claude API model ID to
claude-fable-5-2. On Amazon Bedrock,anthropic.claude-fable-5-2is the base model ID. For cross-Region inference, use the US or global inference profile ID listed by AWS for your routing requirements. - Replay a long session from your application. Keep prior thinking blocks and the preceding system, tools and message history intact. On the Claude API and Google-managed Claude Models on Google Cloud, a supported move from 5.1 to 5.2 can carry earlier thinking blocks forward. A rollback to 5.1 drops thinking generated by 5.2.
- Outside Amazon Bedrock, replace
computer_20251124withcomputer_toolset_20260801. Remove the legacy beta header, omit the old tool name and display dimensions, and update the agent loop for the toolset. If a Fable 5.2 executor uses the beta advisor tool, choose Fable 5.2 or, where you have verified access, Mythos 5.2 as its advisor. Start a fresh session when changing the tool definition. - Rerun quality, latency and cost checks at your intended effort setting. Include multi-turn tool use and requests with a tight output budget, because adaptive thinking consumes part of that budget.
Who Should Upgrade / Who Should Wait
If you already run Fable 5.1 for difficult coding, research or agent jobs, start with a small 5.2 traffic sample. Compare completed tasks, total tokens and elapsed time across the full workflow. If Opus 5.5 meets your quality target, it remains Anthropic's recommended starting point for most work and has half Fable 5.2's standard input and output rates. Schedule the upgrade after testing if your conversation store rewrites prior messages or your agent depends on the older computer-use tool.
