I have been using ChatGPT 5.5 and Codex today, and the difference from 5.4 is immediately noticeable. This is not a careful benchmark, so I do not want to pretend it is more scientific than it is. But for code and logic execution, it feels like a real step up: fewer dead ends, better follow-through, and a stronger sense that the model understands the shape of the task rather than just the next patch.
That lines up with how OpenAI is positioning the release. In its GPT-5.5 announcement, OpenAI says the model is strongest in agentic coding, computer use, knowledge work, and research-style workflows. The Codex developer docs are even more direct about Codex-style work: GPT-5.5 is described as the starting point for demanding agents when available, especially ambiguous multi-step work that needs planning, tool use, validation, and follow-through across a larger context.
The bit that matters to me is not that it can write more code. Models have been able to produce a lot of code for a while. The improvement is that GPT-5.5 seems better at keeping the whole job in its head. It is more willing to inspect the surrounding system, make a plan, edit the right files, run the checks, notice when something did not line up, and keep going. GPT-5.4 was already useful, but it still had more of that stop-start feeling where you had to keep dragging it back to the actual goal.
The NVIDIA write-up is probably the most interesting external example because it is less about a demo and more about operational use. NVIDIA says more than 10,000 employees got early access to GPT-5.5-powered Codex, with engineers using it in complex multi-file codebases and seeing debugging cycles move from days to hours. I would still treat that as launch-post language, but the shape of the claim matches what I saw today: the gain is in persistence and coordination, not just clever snippets.
There is a tradeoff. GPT-5.5 is not just “5.4 but free and faster”. The Codex usage docs say it uses fewer tokens to reach comparable results, but the limits and availability are different, and API access is still its own rollout story. For day-to-day development, though, I care more about wasted loops than raw message counts. A model that takes slightly more careful work per request but gets through the task with fewer retries can be cheaper in the way that actually matters: less time spent supervising nonsense.
My practical take is simple: if GPT-5.5 is available in your Codex picker, I would use it for the hard tasks first. Give it the messy refactor, the confusing bug, the test failure that needs actual reasoning, or the bit of product logic where a shallow answer is worse than no answer. Keep 5.4 or the smaller models around for routine work, but 5.5 already feels like the model I would reach for when I want Codex to carry a thread properly.