Chat with Instant, Thinking and Pro right here. 400K context, adaptive reasoning, and one account that also covers Claude and Gemini.
GPT-5.2 was OpenAI's answer to Gemini 3 Pro, shipped in December 2025 with a much larger context window and adaptive reasoning that scales compute to the question.
Published figures, including the rows where GPT-5.2 doesn't come first. Higher is better.
| Benchmark | GPT-5.2 | Claude Opus 4.6 | Gemini 3 Pro |
|---|---|---|---|
| GPQA Diamond· graduate-level science | 93.2% | — | 91.9% |
| AIME 2025· competition math | 100% | — | 100% |
| SWE-bench Verified· real GitHub issues | 80.0% | 80.8% | 76.2% |
| GDPval· expert knowledge work | 70.9% | — | — |
| FrontierMath· research mathematics | 40.3% | — | — |
| ARC-AGI-1· abstract reasoning | 90%+ | 94.0% | — |
| Context window· input tokens | 400K | 1M | 1M |
Figures from OpenAI's GPT-5.2 announcement and independent benchmark write-ups. The honest picture: Claude Opus 4.6 edges ahead on SWE-bench Verified and ARC-AGI-1, and both rivals offer a larger context window.
Feed it an entire repository or a full documentation set, then get a long-form architecture doc back in a single pass — no chunking, no stitching.
It scales how hard it thinks to how hard the question is, so simple prompts stay fast while genuinely difficult ones get the deliberation they need.
Diff-based edits and context compaction make it well suited to iterative software work rather than regenerating whole files each turn.
Upload a screenshot of a broken dashboard and it reads the error itself instead of asking you to retype it — a real shift for debugging and support work.
Built for agentic workflows through the Responses API, with high accuracy on multi-step tool-use benchmarks and long-range planning.
OpenAI reported meaningful improvements in how GPT-5.2 handles conversations touching on mental health distress compared with GPT-5.1.
The model is identical. What differs is everything around it.
GPT-5.2 ships in three modes, and the difference between them is how much time and compute the model spends before answering.
Instant answers immediately without an explicit reasoning pass. It's the right choice for conversation, quick lookups, rewriting, and anything where you'd rather have a good answer now than a slightly better one in thirty seconds.
Thinking is the sensible default for real work. It reasons through the problem step by step before responding, which matters for debugging, analysis, planning and anything multi-step.
Pro spends substantially more reasoning time and compute than Thinking. Save it for the problems where being wrong is expensive: a tricky proof, a subtle concurrency bug, a decision you'll build on for months.
Claude Opus 4.6 scores 80.8% on SWE-bench Verified against GPT-5.2's 80.0%, and reaches 94.0% on ARC-AGI-1. Both Opus 4.6 and Gemini 3 Pro offer a 1M-token context window. No model wins everything — which is the entire argument for having them all on one account instead of committing to one.