Simular's Sai agent reached 73 percent on OSWorld 2.0 by writing code

Simular reported on 27 August that its computer agent Sai reached a 73 percent success rate on OSWorld 2.0. Adrian Bridgwater covered it for The New Stack.

Every figure in this item comes from Simular, so read them as the company's claims until the leaderboard confirms them. That caveat matters more than usual here, and the article says why below.

The numbers, and where they come from

OSWorld 2.0 is a 108-task benchmark from the XLANG Lab at the University of Hong Kong, launched on 26 June and updated on 8 August. Unlike its first version, tasks are measured in hours rather than minutes: work that takes a skilled person more than an hour.

Simular reports Sai at 73 percent, against GPT-5.6 Sol at 62.57 percent as reported by OpenAI and Opus 5 at 70.57 percent as reported by Anthropic, and says Sai reached the top at roughly two thirds the cost of either.

Note that all three of those figures are self-reported by the company that makes each agent. The Sai results are not yet on the official OSWorld 2.0 site. Asked about that, Simular said it is submitting to the leaderboard and uploading its trajectories to Hugging Face, and pointed out that publishing on your own site first is common practice.

That is a fair answer and it is still a reason to wait for a score somebody other than the vendor produced.

What the tasks look like

OSWorld 2.0 poses things like finding and reasoning across receipts scattered between email and expense reports, handling a message that arrives midway through a task, following reimbursement guidelines precisely, and resolving contradictory data.

That is recognisable work. It is also the kind of work where a small error rate adds up, because nobody checks the hundredth invoice as carefully as the first.

How the cost saving is claimed to work

This is the part worth learning from, whether or not the score is confirmed.

Simular describes a neurosymbolic approach: neural exploration combined with symbolic code, so a solved task can be encoded as reusable code that replays identically every time. Processing the hundredth invoice does not cost what the first one did.

Sai is reported to use roughly 1.5 times fewer model calls on average than pure models, by taking more actions per turn, using a code layer for planning and executing longer subtasks. It keeps input size bounded with adaptive summarisation and holds a constant prompt prefix for as long as possible before summarising, so the start-up cost of each turn is not paid again from scratch. It routes to specialised smaller models for locating interface elements and verification, so the expensive model does not handle steps that do not need full context.

The clearest illustration in the article: Sai, Sol and Opus were all made to play Chrome Dino and reach a score target. Sai treated a repetitive real-time game as a programming problem rather than a clicking one.

The line worth remembering

Simular's chief technology officer Jiachen Yang, on models trained to solve open mathematical conjectures being used to validate invoices: "Posterity will find it ludicrous that people are still building models that way right now."

Source

• "Posterity will find it ludicrous": Sai agent hits 73% on OSWorld 2.0, The New Stack, 27 August 2026