| Vals Index(Google model card; Vals AI index over finance, coding, legal and tax work. Same table: GPT-6 Astra 63.1, Claude Opus 5.5 67.0) | 68.9% |
| Terminal-Bench 4.0(Google model card; Argon self-computed, rivals from the public leaderboard. Same table: GPT-6 Astra 58.2, Claude Opus 5.5 66.4 (Anthropic's own post says 57.9 for Astra)) | 57.4% |
| FrontierSWE v2(Google model card; Proximal public leaderboard, one body for all four models. Same table: GPT-6 Astra 65.5, Claude Opus 5.5 62.3) | 55% |
| DeepSWE v1.1(Google model card; Argon self-computed on a mini-swe agent harness, rivals from the leaderboard and their system cards. Same table: GPT-6 Astra 74.1, Claude Opus 5.5 74.2) | 77.9% |
| Automation-Bench(Google model card; Zapier leaderboard, private set — not AutomationBench-AA 77.5. Same table: GPT-6 Astra 41.4, Claude Opus 5.5 42.5 (Anthropic's own post says 40.0)) | 51.3% |
| OSWorld-2.0(Google model card; offline subset, partial score, best run of 3. Same table: GPT-6 Astra 72.6) | 69.2% |
| Chartography(Google model card; without tools, Surge leaderboard — a different configuration from the tool-assisted entry. Same table: GPT-6 Astra 71.0, Claude Opus 5.5 66.3) | 71.6% |
| GraphWalks(Google model card; 256k-1M band, BFS F1; 99.7 up to 128k. Same table: GPT-6 Astra 71.8, Claude Opus 5.5 66.8) | 84.2% |
| CWE-bench v1(Google model card; public leaderboard, pass@1 with ties broken on pass@4. Same table: GPT-6 Astra 68.0, Claude Opus 5.5 67.0) | 68% |
| Agents' Last Exam(Google model card; binary pass rate, ALE-Claw harness, 5-hour window, safety filters on. Same table: GPT-6 Astra 34.2, Claude Opus 5.5 38.2) | 39.5% |