New SOTA. 6.9pp above Opus 4.6 (72.7%). Desktop agent tasks.

New #1 on OSWorld-Verified, surpassing GPT-5.4 (75%). Achieved with only 10B active params (122B total MoE). Paris-based startup focused on autonomous desktop agents.

Smaller Holo3 variant with only 3B active params. Still surpasses GPT-5.4 on OSWorld at a fraction of the cost.

GPT-5.4 achieves 75% on OSWorld benchmark for desktop tasks, surpassing human-level performance.

Matches human baseline (~72%). Anthropic's best computer-using model.

Sonnet-class model matching human performance level on desktop tasks (humans ~72.4%). Previously only Opus-class models achieved this.

Just below human baseline (72.4%) on desktop automation. Remarkable for a mini-tier model.

Computer-use agent benchmark. Trails GPT-5.4 (75.0%) but competitive for a coding-specialized model. Human baseline ~72%.

Desktop automation score, trailing GPT-5.4 (75.0) and Claude Opus 4.6 (72.7).

Frontier model performance on real-world computer tasks

Research system combining GUI operator (CUA 4o) with coding-as-actions programmer (o4-mini). SOTA on OSWorld at time of publication.

Agent using third-party frontier models. pass@10 reaches 77%, surpassing human baseline (72.4%).

Desktop use agent benchmark. Decent for 27B open-weight model.

Notable computer-use capability for a small/fast tier model.

9B open model beating OpenAI CUA o3 (42.9%) and Claude 4.0 (30.7%) on desktop tasks. Uses ComputerRL framework with API-GUI paradigm.

Desktop automation. Limited by reasoning depth for long-horizon agentic tasks.