When Fable leaves review on Opus PR's, Opus is always impressed.
"The review is excellent — it went through the actual Claude session transcripts from the eval run and identified exactly why the agent went wrong...."
Can't say the same for the other direction.
@simonw The ambitious part is Fable giving me back the attention I used to spend nudging Opus all day.
I’m proud of this project, but I’m not sure I’d call the project itself ambitious
https://t.co/uSLiAdl1Q7
@jasoki@steipete@Steve_Yegge +1 - total catastrophe on anything I used it for.
IMHO it is unusable because the models aren't good enough.
But also, would a good enough model even need GasTown's concepts?
Someone with unlimited tokens should try it with Fable.
@thesaraharminta My guess: they're testing how willing people are to pay per token for higher-quality models.
I can't imagine paying token prices. When it first launched with extra usage enabled, it burned through $50 in seconds.
Hope GPT 5.6 puts some pressure on them, competition is good.
@Terrencemane@atomic_chat_hq It legitimately feels like a good model. I don't do video so I don't know how it does there, for coding it seems certainly Opus class. Perhaps not quite as good as 4.8.
@neetcode1 Workflows via /effort ultracode can burn even Opus limits up fast.
/goal is also pretty good.
Anything that you instruct it to launch many subagents