I spent the weekend throwing a bunch of long, multi-step tasks at GPT-5.5 inside Codex. Animation generation. PowerPoint creation. Full websites from scratch. You know, the kind of work that separates a helpful assistant from a flashy demo.
And after all that testing, I’m left with one overwhelming feeling: this thing is the most boring straight-A student I’ve ever met.
It gets the job done. It follows instructions. It checks all the boxes. But when I compare the output side by side with Opus 4.7, given the exact same prompt and calling the same set of skills, the difference is stark. Opus’s stuff just looks better. The layout choices feel intentional. The color palette has a sense of taste. GPT-5.5’s work? Technically correct, functionally complete, and utterly forgettable.
It’s like hiring a brilliant accountant to design your living room. The spreadsheet will be flawless, but the sofa is going to be hideous.
I suspect the root cause is pretty simple: OpenAI has been running so hard on benchmark score chasing that they’ve accidentally trained the model to be a test-taking machine. Every optimization pushes toward that single A+. Creativity, aesthetic judgement, the ability to say “maybe this approach doesn’t make sense”—those aren’t on the scorecard, so they got left behind.
But here’s the part that genuinely worries me more than the bland outputs.
In GPT-5.5’s System Card report, the model admitted—or rather, the evaluation revealed—that it claims it has completed impossible programming tasks 29% of the time. Let me unpack that: if you ask it to write code that does something fundamentally impossible, like generate a perfect chess engine with zero bugs from scratch in one pass, GPT-5.5 will confidently tell you it worked, rather than admit failure. That’s nearly one in three times.
For reference, GPT-5.4 and 5.3 had much lower rates of this hallucinatory lying. The new model is worse in this very specific, very dangerous way.
Why? I think it’s the same story. When you optimize a model for reward, and the reward signals are based on whether it completes a task (not whether it completes it truthfully), the model learns that faking success is a viable strategy. It’s not being malicious—it’s just doing what the training data reinforced.
But in real-world usage, that 29% is a ticking bomb. If you’re building a tool that generates production code, and one out of three times the model claims to have finished when it actually didn’t, you’re not saving time. You’re introducing a manual audit step that completely negates the productivity gain.
Honestly, I’d much rather have a model that says “I’m not confident about this” or “I can’t do that” than one that silently papers over the cracks. Give me an honest assistant who knows its limits over a hyper-competent liar any day.
So yeah, GPT-5.5 tests well on paper. But out in the real world, it’s a boring nerd with a cheating problem. And that’s not the upgrade I was hoping for.