DevDay 的 Dots 演示翻车与回放
OpenAI 的 Dots 现场演示出故障后放出回放：Astra 驱动的常驻 agent 订到约 326 美元可退机票并筛出两家酒店。
OPENAI'S DOTS DEMO BROKE ON STAGE. THE REPLAY: ONE CALL, AN IN-POLICY REFUNDABLE FLIGHT AT ABOUT $326 AND TWO HOTELS SHORTLISTED The live demo at DevDay didn't go as planned, so OpenAI posted the replay: 2.7M views since yesterday. A dot is an always-on agent with its own computer, running on GPT-6 Astra. In this run it finds a Southwest flight at about $326, pulls hotel screenshots and maps, then works through the user feedback channel. Asked for the biggest theme in that feedback, it names its own: dots that look active but don't reply or can't reach their tools. Astra lists at $10 / $50 per million tokens. The article below moves 12 jobs off it onto GPT-6.1 Sol at $2 / $10, and computer use is one of them. On OSWorld 2.0, OpenAI's own numbers put Sol at 71.4% against Astra's 73.5%, at about $1.30 per completed task against $9.30. An agent that runs all day pays that difference on every task. The leaderboard race went the other way this week. Polymarket has Google at 56% to hold the best AI model at the end of October, up 49 points in a week after Gemini 4 Argon. Anthropic is at 44%, OpenAI at 0.45% http://polymarket.com/event/which-company-has-the-best-ai-model-end-of-october?via=atlas
这条是引用着另一条写的：x.com/0xbobaaa/status/2105273182357762228