Skip to content
AIMODELRACE
Operations

Best agent for browser & computer use

Operating websites and desktop apps that have no API.

Leader

Dominant lead

GPT-6 Astra (ChatGPT agent)

Index

95

Lead

+5.0

% ahead

5.6%

over Claude Opus 5.5.

Full ranking

Task index, 0 to 100. Bars are scaled to the field, so small gaps look small.

  1. 1

    Leads our computer-use race.

  2. 2
    Claude Opus 5.5
    −5.090

    Very reliable multi-step flows.

  3. 3
    Claude Haiku 5.5
    −6.089

    Astonishing value for its size.

  4. 4
    Browserbase + Stagehand
    −10.085

    Infrastructure for browser agents.

  5. 5
    Gemini 4 Argon
    −11.084

    Strong visual grounding.

Key takeaways
  • 1GPT-6 Astra leads computer use; Opus 5.5 and Haiku 5.5 are close.
  • 2Haiku 5.5 is remarkable value for high-volume automation.
  • 3Prefer APIs when they exist: screens are slower and more brittle.

Recommended setup

A computer-use model in a sandboxed browser, with session recording and approval gates before submissions.

Prompts that work

Portal task

Log into {{portal}} with the stored test credentials, download all reports from {{date_range}}, and save them to {{folder}} named {{pattern}}. Screenshot after each download. Stop and ask if anything requires a payment or a new permission.

More in operations