Pine Computer 云电脑 SDK 开启内测
Pine AI 推出面向开发者的云电脑 SDK，用「事件驱动」替代每步重读屏幕，官方称 SaaS-Bench 得分 78.3%、token 成本约为 Opus 5 + Claude Code 的 1/26。
Most computer-use agents are stuck on a problem network servers hit in the late ’90s. After every step, they re-read the whole screen to find what changed. You pay tokens for each of those looks, including the ones that find nothing. Those servers polled every open connection on each pass. The fix was an operating system call that reports only the ready ones. On Linux, that’s epoll. Pine Computer, launched in private beta today by @PineAIAssistant, is a cloud computer developers build on through an SDK. The company calls its design epoll for AI perception. In that design, the AI gets told the moment something changes, so it skips re-reading the screen after every step. By the company’s numbers, the system posted a 78.3% checkpoint score on SaaS-Bench v1.1, highest in the published comparison. Its model-token cost, by the same count, was about 1/26 of Opus 5 + Claude Code. Checkpoint score gives credit for partial progress on a task. Each setup also ran a different model, so I’d read both numbers as whole-system results. Models will get cheaper at reading screenshots, the way faster CPUs made polling hurt less. Servers switched to events anyway. I expect agent infrastructure to replay operating system history, fast. Events replacing polling is one chapter. Three more are already showing up in agent products, this one included. Running many computers in parallel, keeping work alive after the laptop closes, and letting a person take over and hand control back. Operating systems learned long ago that the cheapest check is the one you skip. Agents are relearning it one screenshot at a time.
这条是引用着另一条写的：x.com/1950463458648150016/status/210858888779078…