27B Model Beats GPT-5.5: Weights Unchanged, Interface Changed
Simular's Agent S4 enables Qwen3.8-27B to solve 21.3% of desktop tasks on OSWorld V2 at $1.10 per task, while Opus costs $72.40. The model was not retrained; what changed is the agent's working interface.
The 27B Qwen3.8-27B solved 21.3% of tasks on OSWorld V2. Cost per task was $1.10. Claude Opus 4.8 was $72.40. The tweet says it also beat GPT-5.5. The model was not retrained.
The key point lies at the interface layer. Simular's Agent S4 changes the interface through which the agent works. The project poster's subtitle says it plainly: Agents that discover their own interfaces from experience. The agent discovers its own interfaces from experience.
Retraining means changing the model weights and training again. That was not done here. With the same open 27B model, replacing the outer interface earned more points on desktop tasks. OSWorld V2 tests whether an agent can complete tasks in a graphical interface. Being able to reason does not equal being able to use software. There is a layer of translation between the model and the software. If the interface is hard-coded, no matter how strong the model is, it must follow the same set of actions. If the interface can evolve, a 27B model also has room.
The cost gap is large. $1.10 versus $72.40, about 66×. 21.3% is not high; most desktop tasks still fail. This result shows that the path is worth examining: do not retrain the model; first change the interface between the agent and the software.
Source: Simular tweet
发布时间: 2026-10-25 20:29