definite proof that claude fable subagents silently switch the model to opus even if you have the fallback disabled. absolutely disgraceful treatment of paying customers from anthropic.
Fable 5 AI subagents reportedly route tasks silently to Claude Opus even when fallback settings are disabled
Inspect session logs to verify which model executed tasks
Many users criticized silent switches to Opus in Claude subagents and Fable despite disabled fallbacks because the changes override settings, rapidly burn credits, and feel deceptive.
Most Activity

i sent this to claude and it said it cannot trust delegating work like this.
fable is a smarty model trapped behind anthropic’s insane policies.

@banteg Check out this repo for deeper analytics but yeah true.. https://github.com/TarekAwwad/authrty-claude-code-analytics
FYI if you used fable 5 and built something cool
before you say "wow fable 5 is amazing", check very carefully and scroll through your entire agent session
because it's very possible that it silently switched to opus very early on and it's actually opus who did the job lmao

@banteg It's the way it works. It will route the jobs through a classifier, which may decide to use smaller models for simpler tasks, e.g. I've observed it using Haiku for Explore tasks but falling back to Fable for the actual analysis.

@banteg Yep and you pay for it because of the switch 4.8 you re read your cache

@DavidPrenticeJr it works for reversing, and it knows all the obscure compiler trivia. the refusals are really unfortunate, there is nothing inherently wrong with reversing.

@banteg I really want to use fable for my ongoing RE project, but I can’t with the price, refusals, and stuff like you are pointing out. It’ll only be SOTA for a little while so I am fine waiting.

@kunchenguid Opus gave me shit asking about beans. Fable gave me shit asking about ham.
Somehow, none of them are upset about cornbread.
I thought I'd be slick and ask about cast iron and they banned me

@kunchenguid Lots of ppl keep forgetting that w/ Fable u first handover ur prompt/cmd to a proxy (i.e. classifier).

@kunchenguid You can tell it to show the model in your /statusline, and you will notice it without scrolling to find this message.

@kunchenguid I burned through my weekly in 2 days When it was normally 5 days.
Got it to do security evals after obfuscating my intentions. Found some bugs that even I thought was esoteric.

@banteg 1) sounds like an unanticipated bug, because the "fallback" is a new thing 2) did you expect that disabling it to bypass the fallback? Seems like it, you should be banned for that, wouldn't have to hear you cry anymore if you can't use it

@solipsnitsyn i don't think average users will watch the stats so closely all the time that a silent switch like this will become immediately noticeable

@banteg thank me later

@LefterisJP there is a model field in jsonl rollouts

@banteg how did you mesure it?

@kunchenguid wouldn't you notice because usage limits suddenly 2x'd or are the limits with fable just so low for everyone that nobody can tell?

@banteg Yeah I end up paying a lot when using subagents because of that. It burns through the allowance then through paid credits like nothing.

@banteg There's a simpler way to find this than benchmarking. I actually found it in the first few hours, when I was trying to make a skill for Opus to auto-refine prompts to Fable when a subagent crashes because of rerouting (there's an option in config) and it discovered in json this:

@DavidPrenticeJr claude code