GPT-5.2 (High): benchmark results
GPT-5.2 evaluated at the high reasoning-effort setting. Provider: OpenAI. Released 2025-12-11. Access: API.
Aggregate Rating 1778 ± 6, rank #176 of 2168 rated models, from 382 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| BankerToolBench - Harness - OpenHands | 54.5 | Mean rubric score (%) | 100 |
| CAPITU - Instruction-Level Accuracy | 99.3 | Instruction-level accuracy (%): share of all individual cons | 100 |
| CAPITU - Strict Accuracy | 98.5 | Strict prompt-level accuracy (%): share of prompts whose con | 100 |
| CLEM AdventureGame | 99.17 | Game Clemscore (%), text v3.0; published played × quality / | 100 |
| CLEM Clean Up | 100 | Game Clemscore (%), text v3.0; published played × quality / | 100 |
| CLEM Codenames | 87.69 | Game Clemscore (%), text v3.0; published played × quality / | 100 |
| CLEM Deal or No Deal | 99.12 | Game Clemscore (%), text v3.0; published played × quality / | 100 |
| CLEM GuessWhat | 93.33 | Game Clemscore (%), text v3.0; published played × quality / | 100 |
| CausaLab (3-Node SCMs) | 100 | Reactor frequency accuracy (%; exact prediction of the held- | 100 |
| CausaLab (4-Node SCMs) | 94 | Reactor frequency accuracy (%; exact prediction of the held- | 100 |
| CausaLab (5-Node SCMs) | 92 | Reactor frequency accuracy (%; exact prediction of the held- | 100 |
| CausaLab (6-Node SCMs) | 80 | Reactor frequency accuracy (%; exact prediction of the held- | 100 |
Interactive version: theaggregate.ai/model?slug=gpt-5-2-high · How It Works · Data refreshed daily, snapshot 2026-10-20.