Models
Simo-1 Pro: maximum judgment quality
Simo-1 Pro (simo-1-pro) is for the calls that matter most: high-stakes gating and hard decisions. It scores 96.75% on DecideBench and answers one question over a screenshot in 189 ms.
Why it exists
Some decisions are not routine. An irreversible action, a policy edge case, a near-identical pair of situations that need opposite answers. For those, you want the best judgment available, and you can afford a few extra milliseconds. Simo-1 Pro is that model.
Published latency
Measured on one request over one 1280×720 screenshot, questions answered in the same pass.
| Questions | Simo-1 Pro latency |
|---|---|
| 1 | 189 ms |
| 10 | 324 ms |
| Each added question | ~32 ms |
Published accuracy
Scored with one Choice question per row, using each benchmark’s own options.
| Benchmark | Result |
|---|---|
| DecideBench | 96.75% (pairs 93.5%, easy 98.26%, hard 94.71%) |
| Banking77 | 81.8% (macro-F1 0.813, top-3 94.7%, calibration error 0.047) |
See accuracy for the comparison against other models and how the benchmarks are set up.
Where to use it
- Gating an action you cannot take back, behind a high act threshold.
- Pointing at the exact spot on screen, as in visual inspection and agent supervision.
- Hard moderation and policy calls near the line.
- The second look on cases that Simo-1 scores in the middle band.
Frequently asked questions
How accurate is Simo-1 Pro?
96.75% on DecideBench (94.71% on its hard decisions) and 81.8% on Banking77, with one Choice question per row.
How fast is Simo-1 Pro?
189 ms for one question and 324 ms for ten over one 1280×720 screenshot, about 32 ms per added question.