Nobody Is Competing With Anthropic

Fable 5.2 is faster and cheaper than its predecessor, and it tops Anthropic's own benchmark tables, but in tests on real codebases it trails Opus 5.5 and Sonnet and needs a human close by. Opus 5.5 remains the model to reach for every day, with Fable kept for the hard problems that justify its price [30:20].

Keeping Opus 5.5 as the default and Fable for hard problems

1. Make Opus 5.5 the default model for most coding and knowledge work [07:38].
2. Move up to Fable or Mythos only when the task is ambitious and open-ended, and drop to Sonnet or Haiku when the scope shrinks [07:07].
3. Skip Max reasoning effort for coding, since it costs more and scores worse on most coding tasks [12:10].
4. Enable the design skill for front-end work, because without it the output drifts into generic template styling [32:52].
5. Stay in the loop and audit every output, since Fable makes confident decisions that are obviously wrong to a human [38:40].

Fable 5.2 against Opus 5.5 and the usage limits

Claude Fable 5.2 posts the top score on Anthropic's agentic science benchmark at 74.6, up from 52.6 on Fable 5.1 [08:24]. In the host's own analysis of a TypeScript-to-Rust rewrite, it made 4 false claims out of 61 and ranked behind Opus and Soul. It also generates about 106 tokens per second against Opus's 112 [28:30]. It runs at $8 per million input tokens and $40 per million output tokens [05:47].

Claude Opus 5.5 handles about 63% or more of turns in the host's T3 Code work, the first model to pass 50% of that traffic [18:44]. A $200 subscription that maxes out Opus yields roughly $14,500 to $15,000 in API-equivalent usage, about 73 times the plan price [26:30].

Claude subscription usage limits treat Fable as a separate weekly allowance, and every two points burned there count as one point against the regular limit. Burning Fable alone drops the value multiplier from 73 times to about 15 times [30:12]. A model reset returns half the weekly limit but does not reset the timer [25:11].