/AI2h ago

AI Coding Assistant Silently Downgrades Model Mid-Session on QEMU Port

858646.9K
Original post

Thought I'd found a task that I could use Fable 5 for (porting a QEMU emulator to WASM). But nope, 1.5 hours into the session it tried to disassemble some code to figure out why it was hanging and silently downgraded to Opus 4.8 which has been producing shitty code ever since

8:26 AM · Jun 10, 2026 · 4.4K Views
Sentiment

Positive users enjoy revelations about silent model downgrades in AI coding assistants as they highlight security boundary problems, while negative users criticize the practice as deceptive and unfair to paying customers.

Pos
50.0%
Neg
50.0%
2 comments with sentiment.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS2.3KBOOKMARKS4LIKES30RETWEETS4REPLIES3

I strongly recommend that you set this config option if you're trying to use Claude Code; automatically falling back to a model that's going to write worse code is an insane and unsafe behavior. But personally I do not plan to use Fable 5 until they fix their classifiers.

Thought I'd found a task that I could use Fable 5 for (porting a QEMU emulator to WASM). But nope, 1.5 hours into the session it tried to disassemble some code to figure out why it was hanging and silently downgraded to Opus 4.8 which has been producing shitty code ever since

2hViews 2.3KLikes 30Bookmarks 4

@moyix I'm enjoying watching this, because it reminds me of my long-standing suspicion that there is no boundary between security-sensitive and security-insensitive software engineering. :-)

2hViews 544Likes 7

And if you want to check if any of your sessions were downgraded, the string to search for in ~/.claude is "model_refusal_fallback"

I strongly recommend that you set this config option if you're trying to use Claude Code; automatically falling back to a model that's going to write worse code is an insane and unsafe behavior. But personally I do not plan to use Fable 5 until they fix their classifiers.

57mViews 263Likes 0Bookmarks 0
Tatted@TattedWorks

@moyix The "silently downgraded" part is doing a lot of work here. Anthropic built a product that hides when it stops being the product you paid for, which is either a UX decision or a confession.

2hViews 29Likes 2
papaya ꙮ@papayathreesome

@moyix >until they fix their classifiers i bet in a month itll become more fablous

2hViews 21Likes 2