| Agents' Last Exam | +5.2 points over GPT-6 Astra (new high) | Long-horizon multistep professional workflows |
| Terminal-Bench-Science 0.1 | +9 percentage points over GPT-6 Astra (new high) | Scientific work carried out in a terminal |
| GDPval v2 | +6.2 percentage points over GPT-6 Astra with 22% fewer output tokens | Multistep consulting, finance and design assignments |
| AA-AnalystAgent | 62.5% of questions answered correctly on all 5 independent attempts (new high) | Financial and expenditure analysis over real spreadsheets and reports |
| AA-3D | 1,318 Elo, ranked first (GPT-6.2 Sol: 1,259 Elo) | Blinded human preference on interactive 3D browser builds |
| BenchCAD 2.1 Core | 32% improvement over GPT-6 Astra | Building 3D parts and assemblies from engineering drawings |
| Cybersecurity evaluations | System Card targets: 36.4% solved vs 18.2% for GPT-6 Astra; ExploitBench Internal Port: +22 points over GPT-6 Astra | Separate security evaluations reported by OpenAI |