GPT-5.5 Just Dropped (Yes, Already)
OpenAI shipped GPT-5.5 today. We are barely past GPT-5.4. The release cadence is genuinely getting hard to keep up with, which, honestly, is part of why this site exists.
What's actually new
A few things stand out from the announcement:
- Coding is the headline. OpenAI is positioning 5.5 as their strongest coding model yet, with the biggest gains showing up inside Codex on long-running engineering work — refactors, debugging, multi-file changes, holding context across large systems.
- Same latency, more brain. Per-token latency matches 5.4 but the model is meaningfully smarter. They credit a hardware-software co-design on NVIDIA GB200/GB300 systems and custom scheduling that bumps token generation by ~20%.
- Fewer tokens for the same job. Across their headline evals, 5.5 beats 5.4 while using fewer tokens. That's the opposite of what just happened with Opus 4.7, which has been dragged for being a token guzzler.
- Available now in ChatGPT and Codex for paid subscribers. API access is "coming soon" — they're still wiring up extra cybersecurity guardrails.
Codename was reportedly "Spud," which is somehow the most charming detail in the whole release.
So is it actually better?
Benchmarks say yes. State-of-the-art across 14 evals for generally available models. But you already know how this goes — the benchmark wins and the lived experience don't always match. The question that actually matters is how it feels three days from now, after people have run it through their real workflows.
That's the part benchmarks can't answer.
Help us track it
If you've been using GPT-5.5 today, your gut take is the most useful data we can collect. Head to GPT's page and vote. We just bumped the version label from 5.4 to 5.5, so the score resets to the community's read on the new model from this point forward.
And if 5.5 ends up disappointing, you can always check how Claude and Gemini are scoring right now over on nerfedornot.com.