August 28, 2026 · Research
Sai tops OSWorld 2.0: 73%, at lower cost
Beats GPT-5.6 and Claude Opus on a 108-task benchmark of long tasks that take skilled humans over an hour — and at a lower cost per task.

Our frontier research
Advancing computer-use intelligence. Sharing what we learn.
OUR JOURNEY
August 28, 2026 · Research
Beats GPT-5.6 and Claude Opus on a 108-task benchmark of long tasks that take skilled humans over an hour — and at a lower cost per task.
August 21, 2026 · Product
July 30, 2026 · Research
June 17, 2026 · Research
Argues that agents lack the practice-makes-perfect learning curve, pointing toward a neuro-symbolic compilation approach.
May 19, 2026 · Partnership
Extends Sai into enterprise agent workflows.
May 15, 2026 · Product
“Playwright for the entire desktop”: a desktop automation language built for AI agents.
April 23, 2026 · Product
Lets the agent drive coding tools autonomously — no babysitting required.
February 2026 · Perspective
Lays out the GUI-agent thesis and the neuro-symbolic methodology behind it.
January 26, 2026 · Perspective
A call on the year ahead: desktop agents move from toy to real productivity.
December 15, 2025 · Research
The industry’s first computer-use agent to beat the human baseline (~72%) on OSWorld.
December 2, 2025 · Company
Raised $21.5M to build autonomous computer agents, led by Felicis.
November 18, 2025 · Partnership
Joins Microsoft’s agent-dedicated cloud desktop program.
October 2, 2025 · Research
“Scaling Agents for Computer Use”: replaces deep hierarchies with behavior narratives plus wide scaling.
July 8, 2025 · Product
From research framework to a shippable product.
May 2025 · Research
April 1, 2025 · Research
New SOTA simultaneously on OSWorld, WindowsAgentArena and AndroidWorld.
March 27, 2025 · Partnership
Ties into the Google Cloud ecosystem to accelerate building “the brain of autonomous computing”.
January 2025 · Research
Establishes the computer-use agent framework academically.
December 5, 2024 · Research
Agent S opens up as a pip library for the first time — anyone can run it on their own machine.
October 9, 2024 · Research
The first open agentic framework that uses a computer like a human: experience-augmented hierarchical planning + retrieval-based memory + an Agent-Computer Interface.
August 7, 2023 · Company
Building accessible general intelligence to liberate human labor.
OUR LATEST THINKING
| Date | Category | Title |
|---|---|---|
| Aug 28, 2026 | Research | Sai tops OSWorld 2.0: 73%, at lower cost |
| Aug 21, 2026 | Product | Sai-Fi |
| Jul 30, 2026 | Research | Agent S3 paper accepted to TMLR 2026 |
| Jun 17, 2026 | Research | The Power Law of Practice Is Absent From Agents |
| May 15, 2026 | Product | SimuLang released as open source |
| Feb 2026 | Perspective | The Case for GUI Agents / The Unity of Opposites in Building Computer-Using Agents |
| Jan 26, 2026 | Perspective | 2026: The Year Desktop Agents Stop Being a Toy |
| Dec 15, 2025 | Research | Agent S3 becomes the first to surpass human-level performance: 72.60% on OSWorld |
| Oct 2, 2025 | Research | Agent S3 release + technical paper |
| Jul 8, 2025 | Product | Simular Pro: the first production-ready computer use agent |
| May 2025 | Research | Agent S2 wins Best Paper Award at the ICLR 2025 Agentic AI for Science Workshop |
| Jan 2025 | Research | Agent S paper accepted to ICLR 2025 |
| Dec 5, 2024 | Research | gui-agents v0.1.0 released |
| Oct 9, 2024 | Research | Agent S paper and framework released |
So you're free to think and do, again
Try Sai