Meet your robosecretary
Sai is a robosecretary: a computer use agent that operates a computer the way you do, and keeps operating it after you have stopped watching. It opens applications, reads what is on screen, clicks, types, and files the result where it belongs.
What is a robosecretary?
A robosecretary is a computer use agent that does not stop at running one machine: it runs a whole fleet of autonomous computers on its own. Sai is the first robosecretary in the world.
Sai works a computer the same way a person does: it reads what is on screen, clicks, fills in forms, and moves from one application to the next. Give it an instruction and it goes into the software you already work in and returns the finished task, not a set of steps for doing it yourself.
How reliable is Sai?
Sai ranks at the top of OSWorld, the open benchmark for agents that use real computers. On OSWorld 2.1, the latest release, Sai achieves a 77.3% partial score (79.4% counting repeated runs, 41.7% binary score) at an average cost of $14.34 per task, on par with Claude Opus 5 at 77.67%, whose cost per task is not published. All of Sai's OSWorld 2.1 trajectories are publicly available on Hugging Face.
| Agent | OSWorld 2.1 partial score | Binary score | API cost per task |
|---|---|---|---|
| Sai (Simular) | 77.3% (79.4% with repeated runs) | 41.7% | $14.34 |
| Claude Opus 5 | 77.67% | Not published | Not published |
OSWorld 2.1 results published by Simular on October 2, 2026.
On the previous release, OSWorld 2.0 (108 long-horizon tasks that take a skilled person more than an hour), Sai scored 73.0% at $15.70 per task, ahead of Claude Opus 5 Max Thinking (70.6%, $23.70) and GPT-5.6 Sol Max (62.6%, $26.62). Read the full OSWorld results.
What unifies Sai
Every feature in Sai comes out of the same five commitments.
You watch it work
Sai works inside a live VM View: a window onto the actual machine it is driving, updating as it clicks and types. Steps that are not computer use still report their position in the same place, so a run is never a black box between start and result. You can step in at the point something went wrong instead of auditing the whole output afterwards.
It stops before anything irreversible
Reading, researching and drafting run on their own. Anything with consequences, such as sending an email, making a payment, publishing a post or deleting a record, pauses and waits for a person. Those requests collect in one approval queue instead of interrupting you mid-run, so an unattended task can go right up to the irreversible step and stop there until you say yes.
You describe it, you don't configure it
A recurring task in Sai is a sentence and a cadence: describe the work the way you would brief a colleague, say how often it should happen, and it runs from then on. There are no trigger nodes to pick, no connectors to wire, and no fields to map between steps. Changing it later means rewriting the sentence, not rebuilding the flow.
Any model, including your own
Sai is not a wrapper on one vendor's model. It orchestrates a mixture of frontier and specialist models, routing each step to whichever suits it, and it can be pointed somewhere else entirely: your own API key, an open-weight model, or a model running on your own infrastructure.
One computer, or a hundred
Every account comes with a cloud computer, free plan included, so tasks keep running with your laptop closed. The same work can run on your own machine when it needs your local files or logged-in sessions, or spread across many virtual machines at once: a fleet working in parallel and reporting back to a single queue.
Where to go next
Start with Sai Guide 101, then learn the two features most teams use every week.