Simular Agent S4: The Open Computer-Use Agent That Outscored GPT-5.5 on OSWorld V2
Computer-use agents keep hitting the same wall: feeding screenshots to a multimodal model is expensive, slow, and error-prone. Simular's Agent S4—released October 2026 with results that traveled fast on X—takes a different route called "evolving interfaces." On OSWorld V2, a Qwen3.8-27B-based configuration scored 21.3%, beating configurations built on GPT-5.5 and Claude Opus, at a reported cost of about $1.10 per task. And it is Apache-2.0 on GitHub.
The idea: stop showing the agent raw pixels
Detailed in Simular's Agent S4 article, the framework gives the agent a compact, structured view of the interface—windows, widgets, states, and affordances— rather than a full-resolution screenshot. As the agent acts, the representation is updated to reflect what changed. The agent "sees" less data but understands more of what matters.
The payoff shows up twice in the numbers. First, accuracy: on OSWorld V2—a benchmark of real desktop tasks across real applications—the 21.3% result topped frontier-model configurations. Second, economics: perception tokens are where computer-use bills explode, and smaller representations on a mid-size open model kept cost near $1.10 per task.
Why the model choice matters
The headline configuration pairs Agent S4 with Qwen3.8-27B— a model an order of magnitude smaller than the frontier configs it beat. That inverts the usual computer-use story: instead of paying premium prices for a giant model to tolerate noisy pixels, you pay commodity prices for a focused model working from clean structure. For teams watching costs (and who isn't), open-model computer use at this performance level changes the build-vs-buy math. See also our coverage of DeepSeek V4.1-Flash for where open-model economics is heading generally.
Agent S4 vs the hosted computer-use products
OSWorld V2 scores are for orientation, not procurement. The practical differences between open and hosted computer use:
| Option | Model | Best for | Trade-offs |
|---|---|---|---|
| Agent S4 | Qwen3.8-27B class, evolving interfaces | Self-hosted, auditable, custom-UI automation at ~$1.10/task | You own the runtime, eval, and safety stack |
| OpenAI Operator | Hosted frontier model | Zero-setup web task execution | Closed box; per-task costs and site access vary |
| Claude Computer Use | Hosted Anthropic models | Strong reasoning over GUIs with a documented API | Screenshot-driven perception keeps token bills high |
| Cursor Computer Use | Coding agent with desktop control | Developer workflows—see our Cursor Computer Use explainer | Scoped to dev-tooling contexts, not general UI automation |
Running it safely
An agent that clicks and types on a desktop is an agent that can do damage. Whatever framework you run, pair it with the surrounding stack AIXList tracks:
- kosa8 — a hardware-virtualized microVM per agent, with keys kept on the host, closed networking, and one-second checkpoint/rewind.
- Plango — macOS-native policy gating for agent file access, keys, and network calls.
- CUA Spaces — isolated computer-use environments for agents.
Getting started
Clone github.com/simular-ai/Agent-S, pin a Qwen3.8-27B endpoint (local or hosted), and replay five of your own recurring desktop tasks against it before trusting benchmark deltas. OSWorld V2 is a fine scoreboard, but your internal UIs are the test that pays. Keep humans in the loop for anything touching money, credentials, or irreversible actions.
Agent S4 FAQ
What is Simular Agent S4?
Agent S4 is Simular's open-source computer-use agent framework built around 'evolving interfaces'—instead of raw pixel screenshots, the agent reasons over compact interface representations that it adapts as it works. Released under Apache-2.0 on GitHub (simular-ai/Agent-S).
What did Agent S4 score on OSWorld V2?
In results circulated October 2026, a Qwen3.8-27B-based Agent S4 configuration reached 21.3% on OSWorld V2—above configurations built on GPT-5.5 and Claude Opus—while keeping reported cost around $1.10 per task.
What are 'evolving interfaces' in computer use?
The idea is to give the agent a distilled, structured view of the screen—widgets, states, affordances—rather than full-resolution pixels. The representation evolves as the agent interacts, cutting both the token cost of perception and the error rate of blind clicking.
Agent S4 vs OpenAI Operator vs Claude Computer Use?
Operator and Claude Computer Use are hosted, closed products tuned for consumer workflows; Agent S4 is an open framework you self-host, audit, and extend. If you need reproducible research or custom enterprise UIs, open wins—if you need zero setup, hosted wins.
How much does Agent S4 cost to run?
The framework is free and open source (Apache-2.0). Reported inference cost in the OSWorld V2 runs was about $1.10 per task—your bill depends on the underlying model provider and task mix.
AIXList is an independent AI tools directory. This article summarizes public October 2026 announcements and benchmark results circulated by Simular; we are not affiliated with Simular, OpenAI, or Anthropic. Benchmark configurations change—reproduce results on your own tasks before making architecture decisions.