AI Judge changed title after evaluation, original title: "Claude releases Fable 5, migrating a 50-million-line Ruby codebase in one day during Cognition's evaluation"
Experts debate whether its performance gains over GPT-5.5 are incremental.
No slowing down in sight, this is so weird.
Fable 5 is state-of-the-art on nearly all tested benchmarks, with exceptional performance in software engineering, knowledge work, scientific research, and vision.
The longer and more complex the task, the larger Fable 5’s lead over our other models.
Many users praise Fable 5's benchmark gains and token efficiency over GPT-5.5 for feeling like major progress, while others call the results fabricated PR stunts with questionable real-world value.
While Fable is an amazing model, don't get too excited: it is great, but still has the usual failure models of the other good LLMs we saw in the past, including GPT 5.5. If you look at Anthropic, Opus -> Fable was a huge jump. If you look at the field, GPT 5.5 -> Fable is incremental.
After GPT 5.5 release, OpenAI folks were taking victory laps that they didn't have to nerf a "Mythos class model" and in fact can serve it at scale Well, dust has settled, it's clearly not Mythos class, and Fable is now faster than 5.5. This is embarrassing, but not for Ant
@antirez I find Fable impressive but entirely non enjoyable. It feels like talking to gpt 5 pro. Non interactive and super slow. Probably for people who are running things without being in the loop this works.
I bet right now the maximum you can achieve in the field of software development is to use GPT 5.5 *and* Fable in a combination, cross-review cross-design style. The way I used Opus/GPT when they had similar performances a few months ago.
Sorry, this thing is smarter than 5.5, and less mode collapsed Come on, show what you've got after 1 more month of work
After GPT 5.5 release, OpenAI folks were taking victory laps that they didn't have to nerf a "Mythos class model" and in fact can serve it at scale Well, dust has settled, it's clearly not Mythos class, and Fable is now faster than 5.5. This is embarrassing, but not for Ant
@mitsuhiko Yes basically it does not care about you writing while it works. Looks non-steerable in interactive way in the same way GPT 5.5 is. It's complicated to work like that, especially since right now models are not at a "work alone" level for many tasks.
@francoisfleuret I think of this as the point of inflection where it'll slow down because we'll stop being able to differentiate or measure most improvements to models
bigger models should in theory have more potential capability but the difficulty of the training data is bounded by us
@francoisfleuret It's now impossible to verify these figures because the model is silently steered
@TechMDAI I believe the security thing is not a stunt. GPT 5.5 already uncovered tons of bugs and created a lot of issues in the security community.
@antirez IDK fable has been really wonderful for me, iterally making me rethink my decision of being an engineer lmao
@francoisfleuret No one can prove that biology/cybersecurity bench wasn't fabricated since you can't use it for biology/cybersecurity. It's bullshit.
Welcome to the period after an intelligence leap where it takes a few weeks for people to catch up.
Fable gap up is gpt3.5->4 level at least. The amount of things you can push into it at one time is incredible.
@antirez I totally agree. Its very PR Stunt. The models is held back from the real work or is it really
@francoisfleuret the wall
@antirez they nerfed it though. it thinks everything i'm doing is cybersecurity and/or biology (even though it isn't) and falls back to opus 4.8. this is not really a new model release at all if you can't actually use it.
@antirez I don't think it's isolated at all, but I would bet that we have a k shaped community of people that steer vs people that loop. And the loopers are using more tokens :)
@antirez salvatore, una curiosità, che ne pensi dei modelli open, tipo deepseek v4 pro, rispetto a gpt e fable? Secondo te non sono abbastanza buoni? Per me, deepseek v4 pro funziona benissimo nella maggior parte dei casi (con il prompt/harness/setup giusto).
@antirez Do you think GTP 6 will be the next model ?
It literally told me something like this
"CLANG is not installed on this machine, so i am documenting this in PR body, can't fully test the output"
I was like wtf bro, there is clang installed, test properly and then open PR. It was like Ooooh yeah you are ABSOLUTELY RIGHT
So its same as any other model of clankers
@antirez @yacineMTB As someone who does Blockchain and security work, fable is the worst model of all time.
AI Judge changed title after evaluation, original title: "Claude releases Fable 5, migrating a 50-million-line Ruby codebase in one day during Cognition's evaluation"
Experts debate whether its performance gains over GPT-5.5 are incremental.
No slowing down in sight, this is so weird.
Fable 5 is state-of-the-art on nearly all tested benchmarks, with exceptional performance in software engineering, knowledge work, scientific research, and vision.
The longer and more complex the task, the larger Fable 5’s lead over our other models.
