GPT-5.2
GPT‑5.2 introduces substantial gains in professional knowledge work, outperforming experts on GDPval with 70.9% wins or ties, and setting new highs in coding (SWE‑Bench Pro 55.6%), science (GPQA Diamond ~92–93%), math (AIME 2025: 100%), long‑context accuracy up to 256k tokens, and reliable tool‑calling (Tau2 Telecom 98.7%). It rolls out as Instant, Thinking, and Pro—faster, more structured, and less error‑prone—priced at $1.75/1M input and $14/1M output tokens, with Pro variants supporting xhigh reasoning for top‑quality, end‑to‑end execution.
Benchmark results
| Benchmark | Score | Tags | Source |
|---|---|---|---|
| AIME 2025 | 100.0% | self-reported llm-stats | link → |
| ARC-AGI | 86.2% | self-reported llm-stats | link → |
| ARC-AGI v2 | 52.9% | self-reported llm-stats | link → |
| BrowseComp | 65.8% | self-reported llm-stats | link → |
| BrowseComp Long Context 128k | 92.0% | self-reported llm-stats | link → |
| BrowseComp Long Context 256k | 89.8% | self-reported llm-stats | link → |
| CharXiv-R | 82.1% | self-reported llm-stats | link → |
| FrontierMath | 40.3% | self-reported llm-stats | link → |
| GPQA | 92.4% | self-reported llm-stats | link → |
| Graphwalks BFS <128k | 94.0% | self-reported llm-stats | link → |
| Graphwalks parents <128k | 89.0% | self-reported llm-stats | link → |
| HMMT 2025 | 99.4% | self-reported llm-stats | link → |
| Humanity's Last Exam | 34.5% | self-reported llm-stats | link → |
| MCP Atlas | 60.6% | self-reported llm-stats | link → |
| MMMLU | 89.6% | self-reported llm-stats | link → |
| MMMU-Pro | 79.5% | self-reported llm-stats | link → |
| ScreenSpot Pro | 86.3% | self-reported llm-stats | link → |
| SWE-Bench Verified | 80.0% | self-reported llm-stats | link → |
| SWE-Lancer (IC-Diamond subset) | 74.6% | self-reported llm-stats | link → |
| Tau2 Retail | 82.0% | self-reported llm-stats | link → |
| Tau2 Telecom | 98.7% | self-reported llm-stats | link → |
| Toolathlon | 46.3% | self-reported llm-stats | link → |
| VideoMMMU | 85.9% | self-reported llm-stats | link → |