Reliable
Every step is checked before the next
Independent Judges read the screen and confirm a step worked before Sai moves on. That is how Sai reached 79.38% on OSWorld 2.1, the highest partial score on the official leaderboard.
Compare
vs
Sai API vs Cua
Cua is open-source (MIT) infrastructure for computer-use agents, with local sandboxes, a paid hosted fleet and an optional agent loop, so you assemble the agent yourself. The Sai API gives you the finished agent on a ready computer. Send a task in plain English and get the result back.
- Highest partial score on OSWorld 2.1
- 79.38%
- per task, the only cost published on the leaderboard
- $14.34
- to add Sai to Claude Code, Codex or Cursor
- 1 command
Sai API vs Cua, side by side
As of October 2026
| Capability | Sai API | Cua |
|---|---|---|
| What you get | ||
| Product | A working computer-use agent (Sai) behind one API | Open-source (MIT) infrastructure for computer-use agents plus a paid hosted fleet; not a hosted agent API |
| Computer included | Hosted Windows or Linux cloud computer, or your own Mac or Windows with the Sai desktop app | Yes: local sandboxes and cloud desktops on Cua Fleet (gVisor / KubeVirt) |
| Operating systems | Windows or Linux in the cloud; macOS and Windows on your own machine | Linux, Windows and macOS (macOS local only, on Apple silicon); cloud images Ubuntu 24.04 and Windows Server 2022 |
| Native desktop apps | Yes: desktop apps on Windows, Linux and macOS, not just a browser | Yes: Cua Driver controls native apps on macOS, Windows and Linux |
| Agent loop and action execution | Built in: send a task in plain English, get the result | Optional cua-agent loop (Python, any LiteLLM model); otherwise bring your own agent |
| Setup / time to first task | Create an API key, add one MCP command (about a minute) | pip install cua; first local image pull takes minutes (macOS: 23 GB, ~8 min); you wire up the model |
| Reliability and cost | ||
| Benchmark | Highest partial score on the official OSWorld 2.1 leaderboard: 79.38% (Sai w/ Claude Opus 5, 500 steps, Oct 2026). Binary success 41.67%, #2 behind Claude Opus 5 (44.33%). Trajectories public on Hugging Face | No own OSWorld score; hosts evals (Cua Bench, OSWorld-Verified via HUD) |
| Step verification | Independent Judges check the screen before each next step | Not built in |
| Pricing | Credits per request. Free plan: one shared Windows cloud computer with daily limits. Paid: dedicated computers and top-ups. | Fleet usage: $0.044625 per vCPU-hour + $0.0223125 per GiB-hour; local sandboxes free; model tokens separate |
| Cost per task / cost drivers | $14.34 per OSWorld 2.1 task, the only leaderboard entry with a published cost (OSWorld 2.0: $15.70) | Sandbox vCPU and memory hours (warm pools cost while idle) plus your model spend |
| Repeat runs | Learns a workflow once, then replays it as self-healing code: up to 90% fewer tokens on repetitive tasks | Prompt caching (Anthropic) and trajectory recording; no deterministic workflow replay |
| Control and safety | ||
| Human approval | Built in: approve, approve for this task, or deny; link-only approvals for sign-ins and credentials | Human-in-the-loop model where a person performs steps; no approval workflow |
| Live view / takeover | Live view to watch tasks and answer approvals. No manual takeover | Yes: desktop viewer and Cua Spaces Mac app (watch, take over, share files) |
| Data and compliance | SOC 2 Type II and HIPAA. Runs on Simular's cloud computer or your own machine. | Fully self-hostable (MIT); on-prem available; SOC 2 Type I (May 2026), Type II in progress |
| Operations | ||
| Sessions and persistence | Context carries across tasks on a computer; reset with new-session. | Ephemeral or persistent sandboxes; cloud TTL with keep_alive; no snapshots in SDK 0.9.0 |
| Concurrency and limits | One task per computer at a time; 60 messages per hour; paid plans add dedicated computers | Fleet pools with warm autoscaling; local macOS capped at 2 VMs per Mac |
| Fit in your stack | ||
| Coding agents / MCP | Official MCP server (@simular-ai/sai-mcp) for Claude Code, Codex, Cursor and any stdio MCP client | Yes: Cua Driver runs as an MCP server for Claude Code, Codex and Cursor |
| Model choice | Model-agnostic: Anthropic, OpenAI, Google, MiniMax, xAI, OpenRouter, or auto | Any model via LiteLLM, including local open-weight models |
| SDKs / languages | HTTP API (curl, Python, Node examples) plus the MCP server on npm | Python, TypeScript, Rust, Swift and a CLI |
| Free tier | Yes: free plan includes a shared Windows cloud computer (daily limits) | Open source and free locally; no Fleet free tier documented |
What unifies the Sai API
Four things every task gets, without extra code.
Efficient
Learn once, repeat for less
The first run works the task out. After that, Sai replays it as self-healing code instead of reasoning through every screen, using up to 90% fewer tokens on repetitive tasks.
Ready
A whole computer, already running
Sai works on a hosted Windows or Linux cloud computer, or on your own Mac or Windows machine. It opens apps, clicks and types in desktop software that has no API. No VMs or display drivers to manage.
Composable
Plugs into the agents you already use
Add Sai to Claude Code, Codex or Cursor with one MCP command, or call one HTTP endpoint from any language. It pauses for your approval before sign-ins and other sensitive steps.
Frequently asked questions
What is the difference between the Sai API and Cua?
Cua gives you the parts: sandboxes on your machine or its cloud fleet, Cua Driver for native apps and an optional Python agent loop for any LiteLLM model. You choose the model and wire up the agent. The Sai API is a complete agent: you send a task and Sai does it on a hosted Windows or Linux cloud computer or your own Mac or Windows machine, with step checks and approvals built in.
Is Cua open source?
Yes, Cua is MIT-licensed and fully self-hostable, with an on-premises option. The Sai API is a hosted service; Sai can also run on your own Mac or Windows machine through the Sai desktop app.
Which operating systems does each support?
Cua supports Linux, Windows and macOS: macOS runs locally on Apple silicon only, and cloud images are Ubuntu 24.04 and Windows Server 2022. Sai runs on a Windows or Linux cloud computer, or on your own macOS or Windows machine.
How is Cua priced compared with the Sai API?
Cua’s fleet is billed per resource: $0.044625 per vCPU-hour plus $0.0223125 per GiB-hour of memory, warm pools cost while idle, and model tokens are separate. Local sandboxes are free. The Sai API is billed in credits per request, with the model included, and on OSWorld 2.1 Sai completed tasks at $14.34 each.
Does Cua publish an OSWorld score?
No. Cua hosts evaluations (Cua Bench, and OSWorld-Verified through HUD) but does not publish its own OSWorld score. Sai has the highest partial score on the official OSWorld 2.1 leaderboard, 79.38%.
How long does it take to run a first task?
With Cua you pip install cua, pull a sandbox image (the local macOS image is 23 GB and takes about 8 minutes) and connect a model. With the Sai API you create an API key and add one MCP command, in about a minute.
When is Cua the better choice?
When you want open-source, self-hosted infrastructure, macOS VMs on Apple silicon, native-app control on three operating systems through MCP, any model including local open-weight ones, low raw compute prices or built-in evals. Choose the Sai API when you want a working agent without building the loop, step checks and approvals yourself.
How do I get started with the Sai API?
Sign in at platform.simular.ai and create an API key. Add Sai to Claude Code, Codex or Cursor with one MCP command, or call the HTTP API from any language. The free plan includes a shared Windows cloud computer with daily limits.
Skip the harness. Ship the agent.
Get an API key in a minute, or talk to us about running Sai at scale.