Skip to content

Pine built a dedicated computer for AI: top of the benchmarks, but loses to Codex and Claude Code in real work.

Oct 10, 18:47

Beating AI News Flash: Pine AI launches Pine Computer, a cloud computer designed specifically for AI Agents. It integrates cloud-based virtual machines, AI models, and operational tools. Developers issue tasks via API, and the AI can operate browsers, software, and files in the cloud without needing to use the user's own computer.

Pine's core improvement is enabling the computer to proactively inform the AI about what changes have occurred. Traditional vision-based Agents typically need to take repeated screenshots to determine the results after clicking. Pine, however, has the browser and operating system directly feed back changes to the AI model executing the task, reducing the number of repeated screenshots and page checks.

Pine claims that in internal testing, the new approach achieved task execution speeds 2 to 5 times faster than traditional methods. In the SaaS-Bench v1.1 test, Pine's checkpoint score reached 78.3%, surpassing Claude Code's 74.3% and Codex's 71.1%. This score measures how many checkpoint requirements were completed during task execution, not the overall task success rate.

However, X community notes pointed out that Pine's full task completion rate is only 27.4%, actually lower than Claude Code's 31.1% and Codex's 29.2%. Pine's model costs are indeed lower, averaging about $1.02 per task, compared to $20.50 for Codex and $26.50 for Claude Code. However, these costs do not include cloud server expenses, and the models and execution budgets used by each solution also differ.

Currently, Pine Computer is still in an invite-only beta testing phase.

Source