Models
Reading the situation is the cost. Questions are nearly free.
Simo reads a document, screenshot or video once, and every question runs as a cheap branch off that single read. No tokens are sampled for decisions; the answer is the probability.
Published numbers
Measured on one request over one 1280×720 screenshot, with the questions answered in the same pass.
| Model | 1 question | 10 questions | Per added question |
|---|---|---|---|
| Simo-1 | 64 ms | 104 ms | ~10 ms |
| Simo-1 Pro | 189 ms | 324 ms | ~32 ms |
What the numbers say
- Ten questions on Simo-1 take 104 ms, versus 64 ms for one. The extra nine cost about 40 ms in total.
- Simo-1 Pro takes 324 ms for ten questions versus 189 ms for one.
- So ask everything you want to know in one request. Asking one question per call hides the read-once advantage.
Why it is this fast
There is no reasoning text to write. A reasoning model writes a few thousand words before each answer; Simo answers directly. And because the situation is read only once, additional questions do not pay for the read again.
The result is that a judgment can sit inside the loop: on every agent step, every ticket, every event, every assertion.
Coming next
The full agent step, with decision, argument values, extraction and self-checks, in under half a second on Simo-1 Pro.
Frequently asked questions
How many milliseconds does Simo take?
Simo-1: 64 ms for one question, 104 ms for ten. Simo-1 Pro: 189 ms for one, 324 ms for ten. Measured over one 1280×720 screenshot.
Do more questions make Simo slower?
Only slightly. Each added question costs about 10 ms on Simo-1 and about 32 ms on Simo-1 Pro, because the situation is read once.