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.

OSWorld-Verified variant. New SOTA, narrowly beats Opus 4.7 at 78.0.

New OSWorld SOTA for a generally available model. +5.3 pts over Opus 4.6. 3x higher vision resolution helps computer-use 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.

Open-weight 1T/32B MoE. OSWorld-Verified score. First open model to exceed 73% on OSWorld.

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

Desktop agent benchmark. Tied with Claude Opus 4.6, below Claude Mythos (79.6%) and GPT-5.4 (75.0%).

Agent framework surpassing human baseline (~72%) on OSWorld. Agent S3 alone reaches 66%.

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).

Seed1.8 achieves 61.9% on OSWorld and 85.9% on OnlineMind2Web, surpassing Claude Sonnet 4.5 and GPT-O3-CUA on computer use tasks. Also achieves 70.7% on AndroidWorld.

Seed1.8 achieves 61.9% on OSWorld and 85.9% on OnlineMind2Web, surpassing Claude Sonnet 4.5 and GPT-O3-CUA on computer use tasks. Also achieves 70.7% on AndroidWorld.

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%).

Best open-weight OSWorld result for 32B class. Multi-platform GUI agent.

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

Slightly below Instruct variant on OSWorld. Thinking helps more on web tasks.

Open-source 8B model. Top open GUI agent on OSWorld, competitive with much larger models.

Plan-aware reward model that scores candidate actions for computer-use agents. S3+IntentScore achieves 52.1% on OSWorld, +6.9 over baseline, by overriding nearly half of action decisions.

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.

Plan-aware reward model that scores candidate actions for computer-use agents. S3+IntentScore achieves 52.1% on OSWorld, +6.9 over baseline, by overriding nearly half of action decisions.

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

Seed1.8 achieves 61.9% on OSWorld and 85.9% on OnlineMind2Web, surpassing Claude Sonnet 4.5 and GPT-O3-CUA on computer use tasks. Also achieves 70.7% on AndroidWorld.

Seed1.8 achieves 61.9% on OSWorld and 85.9% on OnlineMind2Web, surpassing Claude Sonnet 4.5 and GPT-O3-CUA on computer use tasks. Also achieves 70.7% on AndroidWorld.

Seed1.8 achieves 61.9% on OSWorld and 85.9% on OnlineMind2Web, surpassing Claude Sonnet 4.5 and GPT-O3-CUA on computer use tasks. Also achieves 70.7% on AndroidWorld.