Claude Fable 5 Review

Fable 5 is the first model that reliably holds a hard problem in its head for an hour without losing the thread. Thinking is not a toggle here, it is always on: the API rejects an explicit disable, and depth is dialed through an effort setting that runs from low all the way to max. That design choice is the whole product. Given a 1M-token window and 128K of output headroom, it will read an entire codebase, build a plan, and execute against it across dozens of tool calls without the drift that makes long agent runs untrustworthy. When it launched in June it sat at the top of the frontier on independent coding-agent evaluation.

Two things keep it from the top grade now. The first is the bill. At $10 and $50 per million tokens it sits at the top of this list's price range, single turns on hard tasks run for many minutes, and it is not frugal with tokens: Anthropic says its own successor, Fable 5.1, does the same token-billed work about 25 percent cheaper and agentic work up to 45 percent cheaper, and by the same company's account Opus 5.5 now matches Fable 5.1 on most work at $4 and $20.

The second is the safety layer. Fable 5 restricts answers in high-risk domains such as cybersecurity and biology, and a request classified as high-risk is handed to Opus 4.8 instead, so security work in particular can quietly drop a generation mid-task. The rest of the operational list still applies: sampling parameters and assistant prefill are gone, it will not run under zero data retention, and prompts tuned for older models actively hurt it.

It launched here as an S, and the reasoning that earned it is still there, which is why critical thinking keeps its S. Graded as it runs today, behind that safety routing and at that price, it is an A.