On the standard Ubuntu desktop benchmark suite OSWorld-Verified, Claude Sonnet 4.6 achieved a 72.5% task success rate, nearly matching flagship Opus 4.6 (72.7%), but still shows some visual localization jitter in high-frequency complex dynamic popups and multi-level right-click menu scenarios.
Model: Claude Sonnet 4.6 (API version claude-sonnet-4-6).
Pricing: Input $3.00 per million tokens, output $15.00 per million tokens.
Benchmark/workflow: OSWorld-Verified (369 cross-application real desktop tasks covering Chrome, LibreOffice Calc/Writer, VSCode, Thunderbird, GIMP, and native OS file management).
Version boundary: Evaluations uniformly run in real virtual desktop containers at 1024×768 resolution, without assistive accessibility label trees (pure pixel screenshot + coordinate click mode).
Scaffold: Standard OSWorld agent loop.
Parameter configuration: effort=medium, max action steps limited to 25 steps/task, mandatory full-screen capture before each action with action trajectory recorded.
Task forms: e.g. "Aggregate sales data by condition in LibreOffice Calc and export as a CSV with a specific name", "Collect product information across multiple Chrome tabs and fill into a purchase form".
| Task category | Sonnet 4.6 success rate | Opus 4.6 success rate | Sonnet 4.5 baseline | Human reference level |
|---|---|---|---|---|
| Full suite aggregate (OSWorld-Verified) | 72.5% | 72.7% | 61.4% | 72.36% |
| Web multi-step forms (Chrome) | 84.2% | 85.0% | 71.3% | 88.5% |
| Office documents and spreadsheets (LibreOffice) | 71.8% | 72.4% | 58.9% | 76.0% |
| File system and system configuration (OS/Terminal) | 79.5% | 78.9% | 69.2% | 82.0% |
| Complex professional software (GIMP/CAD) | 54.5% | 54.8% | 46.2% | 63.0% |
| Average tokens per task | 34.2k | 48.6k | 38.1k | - |
| Average time per task (seconds) | 38.5s | 62.1s | 45.2s | 22.0s |
Cost-performance leap: Sonnet 4.6's overall computer use score differs from Opus 4.6 by less than 0.2%, but average completion time is 38% shorter and token cost is nearly 40% lower, making it highly practical for large-scale RPA and desktop agent deployment.
Exceeding human norm: In standardized, well-structured cross-application data transcription tasks, Sonnet 4.6's success rate already slightly exceeds the normal human tester benchmark (72.5% vs 72.36%).
Failure modes concentrated: Main failures concentrate on: 1) minor pixel-level dropdown arrow click offset (about 35% of failures); 2) action racing ahead due to slow asynchronous network loading; 3) software shortcut conflicts not triggered.
Advanced graphic editing weaker: In GIMP image cropping, layer blending, and other continuous spatial judgment tasks, success rate is only slightly above 50%.
Clone the official evaluation repository https://github.com/xlang-ai/OSWorld.
Configure environment variable ANTHROPIC_API_KEY, and in the config file specify model="claude-sonnet-4-6" and anthropic-beta="computer-use-2026-01-24".
Run the evaluation command: python run.py --benchmark verified --model claude-sonnet-4-6 --max_steps 25.
Export aggregated results and verify artifact file hashes and status using the evaluation script.
Claude Sonnet 4.6
