benchmarks
GPT-5.2 is an OpenAI model for reasoning, coding, image understanding, and long-context work, with newer GPT-5 family models available. The benchmark figures below are OpenAI's published GPT-5.2 Thinking results, not interchangeable measurements of Instant, Pro, or every API reasoning configuration.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| GPQA DiamondPhD-level science questions | 91.4%13 Dec 2025 | View |
| SWE-bench VerifiedResolving real GitHub issues | 73.8%12 Feb 2026 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 96.1%13 Dec 2025 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 64.9%11 Dec 2025 | View |
| FrontierMath | 40.7%13 Dec 2025 | View |
| SimpleQA VerifiedShort factual questions (hallucination resistance) | 37.1%27 Aug 2026 | View |
| LMArena TextHuman preference Elo from blind chat votes (not a percentage) | 1416 Elo22 Oct 2026 | View |
| LMArena WebDevHuman preference Elo for building web apps (not a percentage) | 1417 Elo23 Oct 2026 | View |
Weighted mean of 6 verified benchmark results: GPQA Diamond 91.4%, SWE-bench Verified 73.8%, OTIS Mock AIME 2024-25 96.1%, Terminal-Bench 64.9%, FrontierMath 40.7%, SimpleQA Verified 37.1%.
Capabilities
- Multi-step mathematical and scientific reasoning.
- Software implementation, debugging, and repository-level problem solving.
- Analysis and synthesis across long documents and multiple sources.
- Image understanding, including charts and software interfaces.
- Tool-assisted workflows and production of structured professional deliverables.
What improved
- Higher published scientific-reasoning performance than GPT-5.1 Thinking.
- Stronger published software-engineering results than GPT-5.1 Thinking.
- Improved long-context understanding and integration of dispersed information.
- Improved interpretation of charts and software interfaces.
- Fewer factual errors in OpenAI's evaluated ChatGPT workloads than GPT-5.1 Thinking.
Known flaws & limitations
- Can still generate incorrect facts, reasoning, or code.
- Thinking benchmark results do not establish equivalent performance for Instant, Pro, or every API reasoning setting.
- A large context window does not guarantee complete or accurate use of all supplied information.
- Complex reasoning can increase latency and billed output-token consumption.
Risks, safety & security issues
- Untrusted documents, web pages, and tool responses can introduce prompt-injection risks in connected applications.
- Incorrect tool calls can cause consequential changes when applications grant broad permissions.
- Sensitive information can be exposed through poorly scoped retrieval, tools, or application logging.
- Confidently expressed errors can mislead users when outputs are accepted without independent validation.
Good fit for
- Repository-level coding tasks with tests and human review.
- Scientific and mathematical analysis with independently checked conclusions.
- Long-document synthesis and cross-document research.
- Tool-assisted business workflows with scoped permissions and approval gates.
Poor fit for
- Unsupervised high-stakes medical, legal, or financial decisions.
- Autonomous destructive actions or unrestricted access to sensitive systems.
- Applications requiring guaranteed factual accuracy or exhaustive recall.
- Latency-sensitive tasks where extensive reasoning is unnecessary.
References
Profile last reviewed 10/27/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.