GPT 5.5

OpenAI's previous frontier model, presented at launch as a new class of intelligence for coding and professional work. It remains a current, non-deprecated option for the most complex professional workloads, with configurable reasoning effort and a million-token context. GPT-5.6 Sol supersedes it on most axes but this generation is well understood and widely benchmarked.

GPT-5.5 was OpenAI's frontier model before the GPT-5.6 family, and the pitch was less about raw intelligence than about how little supervision the model needed. The lab described it as understanding what you are trying to do faster and carrying more of the work itself: hand it a messy multi-part task and it plans, uses tools, checks its own output, works through ambiguity and keeps going, instead of waiting to be steered at every step. It remains a current, well-understood option rather than a deprecated one.

On agentic coding OpenAI reported a state-of-the-art 82.7% on Terminal-Bench 2.0 at the time of launch, 58.6% on SWE-Bench Pro, and 73.1% on Expert-SWE, its internal evaluation of long-horizon coding tasks whose median estimated human completion time is twenty hours. It stressed that all three improved on the previous generation while spending fewer tokens, and that the model matched its predecessor's per-token latency in real-world serving despite being larger and more capable. What early testers described was conceptual rather than numerical: understanding why something is failing, where the fix belongs, and what else in the codebase it touches.

The knowledge-work results are the other half of the case. OpenAI reported 84.9% wins or ties on its evaluation of well-specified professional work across 44 occupations, 78.7% on a computer-use benchmark that asks the model to operate real desktop environments unaided, and 98.0% on a customer-service workflow evaluation run without prompt tuning, alongside 60.0% on a finance-agent benchmark and 88.5% on internal investment-banking modelling tasks. On the science side it reported a clear improvement on its genetics and quantitative-biology evaluation, 80.5% on a bioinformatics benchmark, and - with an internal version and a custom harness - a new proof about off-diagonal Ramsey numbers that was later verified in Lean.

Long context was a step change rather than an increment. With a million-token window, the lab reported 74.0% on eight-needle retrieval between 512K and 1M tokens against 36.6% for the previous generation, and 45.4% against 9.4% on the hardest graph-traversal setting at the same depth. That is the capability that makes whole-repository and whole-filing work practical rather than nominally supported.

OpenAI was direct about the risk profile. It treated GPT-5.5's biological, chemical and cybersecurity capabilities as High under its Preparedness Framework, said the model did not reach the Critical cybersecurity level but was a clear step up on cyber over the previous generation, and shipped stricter cyber classifiers while warning that some users would find them annoying at first. Verified defenders could apply for less restricted access. All of the published evaluations were run at the xhigh effort setting in a research environment, which is worth remembering when comparing them with anything measured at a default.