tfw none of my Fable5 chats has been downgraded...
Meta researcher Lucas Beyer says his Fable5 chats haven't been downgraded, sparking debate over silent backend model routing
Story Overview
Meta researcher Lucas Beyer noted steady results across his Fable 5 sessions and shared the observation with a distressed Spongebob meme, yet the comment prompted immediate pushback on whether Anthropic's backend might quietly substitute other models without alerting users.
Routing Transparency Remains Unclear
Florian Brand replied that unchanged performance could result from undetected switches to fallback models, leaving open whether Anthropic's stated notification policy holds in every case.
Everyday Sessions May Hide Complexity
The exchange shows how even experienced users struggle to verify which model actually handles their prompts when routing decisions stay invisible.
Positive users view no downgrades in Fable5 chats as a good sign for production readiness while negative users report frequent downgrades and suspect selective whitelisting to mislead observers.
No Digg Deeper questions have been answered for this story yet.
Most Activity
@giffmana You sure your chats don’t just get routed?
tfw none of my Fable5 chats has been downgraded...
@xeophon Yeah unless they lie to me in the ui
@giffmana You sure your chats don’t just get routed?

@giffmana S-tier move of Dario to whitelist your account to trick you into posting about how the downgrade rate were not nearly as obnoxious as people think

@giffmana I got downgraded for asking it to create an html artifact of a GitHub repo, but Fable did not have any issues discussing building control plane software for a Huawei Ascend SuperPoD I “found in the dumpster with its cooling units”

@xeophon @giffmana You can literally turn that automatic switching off

@xeophon And it literally told me stuff like "that'd be a first. Definitely paper material" lol

@giffmana The only time I got downgraded is when I asked some JavaScript question lol

@Jaidcel @giffmana

@Jaidcel He playing 4d chess with me and i don't even realize. Such a smart fella!

@giffmana Same. Happened twice before, quite unnecessarily. Guardrails working better IMHO

@giffmana ngmi but you already made it

@giffmana Genuinely feels like the classifiers are less restrictive but more random

@giffmana uhmm that’s lucky, mine was working and had the idea to deploy a tunnel so that I could reach the demo, and it instantly fall to 4.8

@giffmana can't say the same. even when i tell it to avoid checking security in a performance audit, it downgrades :P

@giffmana probably nothing — but if the cap holds across the board, that’s a signal the team is running tighter than expected.

@julianw1011 @giffmana I know, I’ve set it to off. Switching is on by default, though. And it’s in like 3 different places

@giffmana Interesting—wonder if that means the eval criteria shifted or your agent’s context window just handled it better. Either way, good sign for production readiness.

@giffmana

@giffmana weird i cant talk to it about anything

@giffmana I'm being hit hard today: using ephemeral token for smoke authentication and getting ui screenshot is fatal, instant opus 4.8 all the time.