Efficient execution
Adapts experiments to available hardware and uses half the tokens of direct OpenAI prompting.
A research engine
OpenAssay develops a research plan from a single prompt and executes it end to end, with iterative evaluation and self-correction.
In development · Public release coming soon
Adapts experiments to available hardware and uses half the tokens of direct OpenAI prompting.
Links sources, plans, experiments, results, and corrections in one research record.
Papers from 2026 A* conferences, reproduced with accompanying code.
All code and papers will be released.
| System | GPU hours per paper ↓ |
|---|---|
| OpenAssay | 1,300 |
| OpenAI Sol 6.1 | 2,600 |
| Opus 5.5 | 2,400 |
GPU-hour measurements use the same H200 SXM hardware.
50% fewer GPU hours than Sol 6.1; 45.8% fewer than Opus 5.5.
All code and papers will be released. Register to receive updates on OpenAssay and the reproduction collection.