Asana cuts model costs 76x in browser tests with GPT-6.1 Sol (original report, opens in a new tab)
Read the original report from OpenAI.
Inspect the documented capabilities, access conditions and sources for your task.
Initial research: 2026-09-21. Check current terms before choosing.
AI extraction with a second evidence validation pass. These dates apply to the listed fields.
Updated
Browser-scoped rather than full desktop control
Source for Computer use (opens in a new tab)Opening pages, clicking buttons, typing and filling out forms, driven by any LLM the developer chooses
Source for Browser control (opens in a new tab)bu-2-0 lifted internal-eval accuracy from 74.7% to 83.3%
Source for Multi-step autonomy (opens in a new tab)Runs wherever the developer hosts it; Browserbase and Skyvern are the hosted-infrastructure alternatives for enterprise workloads
Source for Sandboxed execution (opens in a new tab)Automatic source checks. These do not change the verification date of the facts above.
Read the original report from OpenAI.
Beyond feature support
How well a model performs depends on the task, version, and settings.
We have not published a reviewed quality evaluation for these versions yet. Feature support and price do not establish output quality.
Try the same small task with each option. Check accuracy, consistency, and how much editing you need to get a usable result.