Screen-level OS controlOpen Source#3 of 14 in category
UI-TARS-2
ByteDance · Launched Sep 2025
ByteDance's end-to-end GUI-grounded agent — UI elements as first-class objects, not pixels to guess. OSWorld-Verified ~53%, AndroidWorld 73.3%, Online-Mind2Web 88.2%.
Visit UI-TARS-2 →Open weights
3
Benchmarks scored
88.2
Peak score
0
Article mentions
Yes
Open source
Benchmark performance
Online-Mind2Web
Live-web successor to Mind2Web with a programmatic checker instead of an LLM judge - much harder to game.
88.2
★ SOTA
AndroidWorld
Google's 116 hand-built tasks across 20 real Android apps on a live emulator. Mobile is harder than desktop for most agents.
73.3
OSWorld-Verified
369 real tasks on a live Ubuntu desktop VM: file I/O, spreadsheets, creative apps, settings. The July 2025 Verified rebuild moved to AWS (50x parallel) and fixed 300+ task bugs. The flagship computer-use benchmark.
53.1
Other screen-level os control agents
The 14 agents in this category, ranked by peak benchmark.
| Agent | Maker | Launch | Peak | Pricing |
|---|---|---|---|---|
| Claude Sonnet 4.6 | Anthropic | 2026-02 | 1470.0 | $3 / $15 per M tokens |
| Claude Opus 4.8 | Anthropic | 2026-05 | 88.6 | $5 / $25 per M tokens |
| Claude Opus 4.7 | Anthropic | 2026-03 | 87.6 | $5 / $25 per M tokens |
| Claude Mythos Preview | Anthropic | 2026-04 | 86.9 | Research preview |
| Gemini 3.1 Pro | Google DeepMind | 2026-02 | 85.9 | Google API |
| GPT-5.5 | OpenAI | 2026-05 | 82.6 | OpenAI API |
| Holo3-35B-A3B | H Company | 2026-04 | 82.6 | H Company |
| GUI-Owl-1.5OSS | Alibaba | 2026-02 | 80.3 | Open weights |
| Holo3-122B-A10B | H Company | 2026-04 | 78.8 | H Company |
| GPT-5.4 | OpenAI | 2026-03 | 75.0 | OpenAI API |
Quick facts
- Type
- Screen-level OS control
- Maker
- ByteDance
- Launch
- 2025-09-01
- Open source
- Yes
- Pricing
- Open weights
- Benchmarks scored
- 3
- Article mentions
- 0
- Rank in category
- #3 of 14