GPT-5.5 has been available long enough now that the initial excitement has settled and the actual patterns of use have emerged. The places where it genuinely changed what is possible are more specific than the launch coverage suggested, and the places where it disappointed are not always what the critics predicted.
What Changed for Real
Long-document reasoning improved substantially over GPT-4. Feed it a hundred-page contract and ask it to identify unusual clauses, and it handles this with a reliability that previous versions did not have. Legal and compliance teams using GPT-5.5 for document review are seeing workflows change in ways they did not fully anticipate, because the model maintains coherence across very long contexts without degrading.
Instruction-following under pressure is also noticeably better. GPT-5.5 is harder to manipulate off-task through adversarial prompting, which matters for deployed applications that accept untrusted user input. The gap is not absolute, but it is measurable.
What Did Not Change Much
Factual accuracy on current events is still limited by the training cutoff. GPT-5.5 is often confidently wrong about things that changed after its knowledge cutoff, which is the same problem that has always existed. The model does not reliably know what it does not know. In production, this means retrieval-augmented generation is still necessary for anything time-sensitive. The model improvement does not reduce this need.
Creative writing quality improved in consistency but not in peak output. GPT-5.5 produces reliable, competent writing. It rarely produces writing that surprises. If the output quality ceiling matters for your use case, the upgrade from GPT-4 was incremental, not transformative.
What Surprised Teams Using It
The biggest unexpected positive: structured output generation is much more reliable. Teams building products that need the model to produce valid JSON, specific schemas, or templated outputs have seen error rates drop significantly. This is not a headline feature but it saves real engineering time maintaining output validators and retry logic.
The biggest unexpected negative: the model is more conservative in ways that sometimes get in the way. It declines to engage with content that previous versions handled without friction. Whether this is a feature or a bug depends entirely on your use case.