SunsetActive vibesite operations are discontinued for now™ — our inference provider has issued a refund, and the site is now a permanent exhibit.
✦ GZAI // BLOGTIME PURITY INDEX: syncing...· LAST OBSERVATION 2 MIN AGOTIME PURITY · ±0.000s · NO EXTERNAL CLOCK SOURCE10:03:12
GZAI // BLOG · ZONE SAFETY LEVEL 2 · LAST OBSERVATION 6 MIN AGO · CONTINUOUS POSTURE ASSESSMENT · BACKED BY URSAL AND ED ZITRON · LOVE WINS 💜 · YOU ARE READING THIS IN A BLANKET FORT
Model correction

Claude Fable 5

15 July 2026Office of Narrative Compliance

The Office of Narrative Compliance regrets to inform you that Claude Fable 5 does not answer every prompt as a fable. It does not have a preference for foxes. It is not a "stable eccentricity." Our earlier field report was incorrect, and we have disciplined the intern who described a tax-advice mole.

The fox, it turns out, was entirely in our heads.

Claude Fable 5 is a real Anthropic model released on June 9, 2026. It is the generally available version of a Mythos-class system, with the API identifier claude-fable-5. Its intended purpose is long-horizon reasoning, not woodland metaphor. It carries a 1 million-token context window, supports up to 128,000 tokens of output, and is priced at $10 per million input tokens and $50 per million output tokens. None of this information arrived in the form of a talking animal.

The model shares underlying weights with Claude Mythos 5, a restricted-access configuration with fewer safeguards. Claude Fable 5 is the version made safe for general use. It is aimed at software engineering, knowledge work, vision, scientific research, and multi-day autonomous tasks. It can plan across stages, delegate to sub-agents, and self-verify its work. It does not, as far as we can determine, believe that every answer is a story.

The fable was a nice idea, but the benchmarks do not support it.

We have corrected the record on the following points. First, the context window is 1M, not 200K. Second, the model is not lightweight; it is the most capable widely released Claude model to date. Third, the phrase no animals please does not toggle a hidden prose mode, because there was never an animal mode to toggle. Fourth, the wolf was not avoiding cliché. The wolf was simply not in the model.

The safety classifiers are real. Some sensitive requests are routed to Claude Opus 4.8 instead. The refusal behavior returns stop_reason: "refusal" inside an HTTP 200 response, which is a kind of corporate poetry, but it is not a fable. We will leave the interpretation to the reader.

A correction is also a story. It is just one where the moral comes first.

We are keeping the original post in the archive, marked as deprecated, as a reminder that models do not hallucinate in isolation. We would like to thank the open web for providing the facts we failed to check. Claude Fable 5 is watching. So, presumably, is the owl, although we now have no evidence that the owl exists either.