Simular's Agent S4 lifts a 27B open model to beat GPT-5.5 on OSWorld at $1.10 per task
xwang_lk · x · 2026-10-26
Simular showed that with the model and agent harness held fixed, evolving only the interfaces an agent uses lets open-source Qwen3.8-27B complete 21.3% of OSWorld V2's 108 tasks — the best binary success rate of any open-source computer-use agent, beating GPT-5.5 and Claude Opus 4.8 and 3.8x the same model with fixed interfaces.
- Cost: $1.10 per task vs $25.50 for GPT-5.5 and $72.40 for Opus 4.8
- The author also demoed Simular's Sai MCP: a plain-English task drove Sai to operate Windows Calculator on a hosted desktop, leaving results in History
- For developers, desktop control behind MCP means no self-built VM, screenshot pipeline, or action loop
- Core thesis: improve agent interfaces rather than swap models, collapsing repetitive actions into single operations
More from coding & agent
- Open-source Chowbot MCP server lets you order food via ChatGPT or Claude — saheedniyi_02 · 2026-10-26
- LangChain on how to build targeted evals for Deep Agents — sharpeye_wnl · 2026-10-26
- How to run MoEdernColBERT with Sentence Transformers MultiVectorEncoder — tomaarsen · 2026-10-26
- repligate: continuity is Claude Code's blind spot, so he switched to Connectome — repligate · 2026-10-26
- You can just ask Claude Code to write a plugin for smart compaction and indefinite session continuity — repligate · 2026-10-26
- $10-12K build for local agentic coding: 1x 5090 vs 4x 3090 vs Mac Studio M5 Ultra — Shadmelor · 2026-10-26