Claude Fable 5.2 ships with 1M context, 30% faster output and a lower price
Anthropic released Claude Fable 5.2, its most capable model for open-ended technical work: same million-token context, output up to 30% faster and a lower price per token than the previous version. The practical question isn't whether it's better, it's when it's worth paying for.
What Claude Fable 5.2 brings
The model landed on November 9 with a clear spec sheet: a 1,000,000-token window, up to 128,000 output tokens and a June 2026 knowledge cutoff. Adaptive extended thinking is always on, with effort levels from low to max — the default is high — and preserved thinking across turns, so a long session doesn't restart from scratch on every message.
Pricing dropped compared with Fable 5.1: $8 per million input tokens and $40 per million output tokens. Context caching runs at $10 per million for 5-minute writes, $16 for one-hour writes and $0.20 per million for reads. The Batch API keeps its 50% discount.
Fast mode and the Mythos variant
There's an optional fast mode that speeds token generation up to 2.5 times, at $16 per million input and $80 per million output. It's for when the clock matters, not for batch processing. Fable 5.2 and Mythos 5.2 are also the same base model: Mythos is the variant with fewer safeguards, available only to organizations verified through the cybersecurity and life sciences programs.
Read also
The numbers that drive the decision
In Anthropic's table, Fable 5.2 sits above Opus 5.5 in almost everything evaluated. On FrontierCode 2.0 it scores 49.1% against Opus 5.5's 45.5%; on FrontierSWE v2, 69.9% against 62.3%. Computer use (OSWorld 2.1) reaches 89.3% and agentic search (WANDR) 82.1%. The comparison rows include GPT-6 Astra and Gemini 4 Carbon, and the model leads most of them.
Benchmarks don't pay the bill, so read them with the vendor's own caveat: for most coding and knowledge work, Anthropic recommends starting with Opus 5.5, which is cheaper ($4 and $20 per million), and moving up to Fable only when the problem demands it. Dropping to Sonnet 5.5 or Haiku 5.5 is also a valid move.
The advisor and orchestrator pattern
The most useful part of the announcement isn't the score, it's how they suggest using it. Fable 5.2 can act as an advisor: a small or mid-size model consults it when stuck instead of forwarding all the work. Or as an orchestrator: it delegates simple tasks to cheaper models and keeps the hard ones. Either way total cost drops because most tokens never touch the expensive model.
How to test it without burning budget
The model is available in the API as claude-fable-5-2, and also on Amazon Bedrock, Google Cloud and Microsoft Foundry. As of November 9 it's also in GitHub Copilot, where Anthropic reports it solved more tasks than Fable 5.1 while using fewer steps and fewer tool calls.
A reasonable plan for this week: pick three real tasks from your repository, the ones that currently take two or three attempts, and run them with Opus 5.5 and with Fable 5.2 while measuring tokens and time to an acceptable result. Compare cost per finished task, not cost per token. With caching on and effort set high, the difference usually shows up where the model avoids a full correction cycle.
Conclusion
Fable 5.2 is the model for work that doesn't fit in a single pass: large migrations, agent sessions that run for hours, research that ends in a finished document. For everyday work, Opus 5.5 still makes more sense, and the real savings come from combining them well. If you want model comparisons judged by engineering criteria rather than headlines, this blog keeps publishing the numbers.
