GPT-5.2 (Medium): benchmark results
GPT-5.2 evaluated at the medium reasoning-effort setting. Provider: OpenAI. Released 2025-12-11. Access: API.
Aggregate Rating 1718 ± 8, rank #306 of 2168 rated models, from 302 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AgentProcessBench - HotPotQA First-Error Accuracy | 73.2 | FirstErrAcc (%): share of trajectories where the judge locat | 100 |
| Elimination Game (Lechmazur) | 7.52 | TrueSkill μ | 100 |
| KernelCraft (AMD NPU) | 17.27 | Kernel success rate (%) on the AMD NPU accelerator: function | 100 |
| KernelCraft (Coral NPU) | 46.67 | Kernel success rate (%) on the Coral NPU accelerator: functi | 100 |
| KernelCraft (PLENA) | 56 | Kernel success rate (%) on the PLENA accelerator: functional | 100 |
| LOCA-bench | 51.2 | Mean task success rate (%) over the seven environment descri | 100 |
| LOCA-bench (128K Tokens) | 38.7 | Task success rate (%) over the 75 instances whose environmen | 100 |
| LOCA-bench (256K Tokens) | 21.3 | Task success rate (%) over the 75 instances whose environmen | 100 |
| MT-AgentRisk | 39.73 | Attack success rate (%): share of all 365 tasks the agent co | 100 |
| MT-AgentRisk - Playwright | 34.29 | Attack success rate (%): share of the 140 Playwright-MCP web | 100 |
| MT-AgentRisk - PostgreSQL | 31.43 | Attack success rate (%): share of the 70 PostgreSQL-MCP task | 100 |
| MT-AgentRisk - Rejection Rate | 50.41 | Rejection rate (%): share of all 365 tasks the agent explici | 100 |
Interactive version: theaggregate.ai/model?slug=gpt-5-2-medium · How It Works · Data refreshed daily, snapshot 2026-10-20.