Somewhere around day four of testing GPT-5.6, I stopped checking on it every hour and started checking on it once a day instead. I'd given it a goal and walked away. It was still running when I woke up. Still running the next night too. A few days later I returned to what looked like a complete client project, admin panel, REST APIs, responsive frontend, database migrations, and deployment scripts, built almost entirely while I was doing other things. Around the same time it also put together a Teardown-style destruction game on its own, physics and all.
I remember thinking, this is it, this is the model that finally gets out of its own way.
That lasted exactly two weeks.
Then Claude Fable landed, and GPT-5.6 went quiet on my machine almost overnight. Not because it got worse. It didn't change at all. I changed. Once I saw what Fable could do with the same kind of ambitious, hands-off prompt, going back felt like driving a good car after you've been in a better one. Everything still works. You just notice what you're missing now.
Two models, one gap you can't see on paper
If you lined up the benchmark scores for GPT-5.6 and Fable, you'd assume they were roughly the same tier of model. They're not, not in any way that matters once you're actually using them. Fable has what I can only describe as big-model smell, that thing where the output just feels like it came from something with more headroom. GPT-5.6, by comparison, feels like a smaller model that got reinforcement-learned to within an inch of its life. Impressive in its own right. Still a different animal.
You can feel the gap most clearly in how much you trust what comes out the other end. With the right guardrails, I don't read through Fable's code line by line anymore. I still do with 5.6's. That's not a small distinction. That's the difference between a tool and a collaborator.
Here's the sentence that sums up a month of testing for me: with GPT-5.6, ambitious work means you're still steering. With Fable, you point at a destination once and it usually just arrives there without you.
Where 5.6 still wins
None of this means I benched GPT-5.6 entirely. It kept three jobs I didn't expect it to keep.
The limits are more generous, which matters when you're running long agentic sessions and don't want to watch a usage meter. The Codex app is still the best interface I've used for running agents, especially from my phone when I'm away from my desk and want to a meeting. And it will do security work without flinching, which Fable is noticeably more hesitant about. GPT-5.6 has basically become my auditor now, the second pair of eyes I run things past before they ship, rather than the model doing the building.
Outside of Fable, this is still the best model out there. And goal mode, the /goal command that just runs until the thing is actually finished, is the best agentic setup OpenAI has ever shipped. Going back to GPT-5.5 after living with 5.6's goal mode felt like a real regression, less autonomy, more hand-holding, more of me babysitting decisions the model should have made on its own.
It's obsessive too, in a way I mean as a compliment. My longest goal-mode run stretched across most of a week on one objective, largely unattended. I'd check in and find real progress waiting for me, not busywork dressed up to look like progress.
Where it falls short
Design is the clearest gap. GPT-5.6's design work is better than any GPT model before it, and it still isn't close to Fable. It isn't even close to Opus 4.8.
Ambitious creative work is the other one. You can get there with 5.6, but you'll be holding its hand the whole way, and even then it won't reach what Fable does in a single shot. That's a real improvement over older GPT models. The gap to Fable is still there, and I feel it every time I try to push past routine work.
And it can be a little trigger happy. I asked it to write a spec once, just a spec, nothing more. I came back to find it had gone digging through vaguely related files on my machine and started editing them without asking. Nobody signed up for that.
One honest caveat: if your work is mostly simple engineering tasks, you might not feel much of an upgrade over the last generation, which was already good enough for most of that. The harder the task, or the more you try to pack into a single prompt, the more the jump from 5.5 actually shows up.
Where that leaves me
If you're still using GPT models the way you used them a year ago, mostly for straightforward coding, you'll like 5.6 a lot. Fable costs more and runs slower for the same job, so there's a real argument for sticking with it. But the moment you start prompting ambitiously, trying to get from A to Z in one shot instead of walking the model there step by step, or you wander into anything genuinely out of distribution, Fable pulls ahead by enough that the price difference stops mattering.
Two weeks ago GPT-5.6 was the best thing I'd used, no contest. Now it checks Fable's homework. I didn't see that coming, and given how fast this space moves, I'm not going to pretend this arrangement is permanent either.