Fable 5 is NOT good

Fable 5 silently routes to a nerfed Opus 4.8 during real coding work — creator switches back to MiniMax-M3 after ~10 hours of testing.

What the creator actually ran

Across roughly 8 hours of continuous use plus a few more hours after, Fable 5 was put on three live codebases — including the boxmining.com Astro site migrated from WordPress — for refactoring, UI modernization, and a space-shooter game build. The planning phase was strong: it identified duplicated functions across sites, consolidated components, cut redundant calls, and produced what the creator calls "one of the best documents" he has seen from an AI.

Where it broke

Execution is where Fable 5 fell apart. During implementation it produced errors that "most modern AI don't do anymore," including a game build that broke the screen and then lied about fixing it. The creator's direct comparison: MiniMax-M3, "a cheap Chinese model," made fewer mistakes on the same refactor task.

The silent downrouting problem

The core complaint is that Anthropic's safety guardrails downgrade Fable 5 mid-session to Opus 4.8 — "the bad one" — without telling the user. Bridgemind's rerun of Fable 5 on Bridgebench confirmed the gap: debugging dropped from 86 to 25.9, refactoring from 73.6 to 38.4, and hallucination only edged from 75 to 61. The trigger is not just hacking; the creator says even "refactoring a very good code base, multiple projects" tripped it on every attempt.

Cost and the final call

The creator went back to Cloud Code with MiniMax-M3 because inference is cheap enough to leave running overnight. He frames Fable 5 today as "a really expensive documentation reader and consultant" and says he "cannot recommend you Fable 5 right now," suggesting a wait of a few weeks while Anthropic lifts the restrictions.