| Ben | Ben FlashFlash | Gemini 4 Carbon | Claude Opus 5.5 | GPT-6.2 Astra | |
|---|---|---|---|---|---|
| FrontierCode-AA 2.0FrontierCode | 35.9% | 40.2% | 40.6% | 46.4% | 45.4% |
| Terminal-Bench 4.0Terminal-Bench | 56.6% | 58.6% | 65.2% | 59.6% | 64.1% |
| Terminal-Bench-Science 0.1Terminal-Bench-Science | 49% | 67.6% | 64.8% | 59% | 78.1% |
| GDP.pdf | 28.8% | 28.8% | 28.4% | 26.2% | 29.8% |
| AA-Briefcase v1.1AA-Briefcase | 1,682 | 1,541 | 1,611 | 1,800 | 1,644 |
| GDPval-AA v2.1GDPval | 1,709 | – | 1,733 | 1,866 | – |
| AutomationBench-AAAutomationBench | 63.2% | 69.9% | 77.8% | 69.5% | 77.2% |
| HLE-Diamond | 55.6% | 58.3% | 69.1% | 67.9% | 72.2% |
| CritPt v2 (Hard Subset)CritPt | 59.5% | 61.5% | 69% | 68% | 71.5% |
| AA-Omniscience v1.1AA-Omniscience | 42.4 | 38.3 | 53.5 | 48.1 | 50 |
13% of the work ran on split threads
Medium and Ben Flash score higher for fewer tokens a task, in about half the time.
Medium scores higher for fewer tokens a task.
The first release: three levels, from quick answers to Ben’s strongest work.
Ben 1.2 is live in Ben today, at High unless you turn the dial. See the plans →
Benchmarks: FrontierCode-AA 2.0 · Terminal-Bench 4.0 · Terminal-Bench-Science 0.1 · GDP.pdf · AA-Briefcase v1.1 · GDPval-AA v2.1 · AutomationBench-AA · HLE-Diamond · CritPt v2 (Hard Subset) · AA-Omniscience v1.1. Scores, cost and time a task as Artificial Analysis ran them, 20 Nov 2026. Cost is each model's list price a task against what Ben High's tokens are worth.