9月25日AI快讯：ChatGPT Pro Max定价曝光，Gemini 3.8 Live正式上线
一天的大事都在这：OpenAI 500 美元档订阅、Gemini 3.8 视频智能体、Perplexity 新检索栈快了一倍多，几分钟扫完全部。
OpenAI 的 ChatGPT Pro Max 定价为每月 500 美元（含 VAT 为 600 美元），新增 Fastest Work 和 Codex 权益，Ultra Fast 路径可能由 Cerebras 支持。Google 的 Gemini 3.8 Live Avatar 在 Gemini Enterprise 正式上线，支持 97 种语言的唇形同步视频智能体，TPU 原型已随 SpaceX Transporter-18 进入轨道测试。Anthropic 恢复对 Claude 触发安全拦截的请求收费，99.7% 的 Claude Code 账户未触发新增计费拦截。Perplexity 在 Photon Rust 检索栈上推出 Fast Search，95% 结果在 230 毫秒内返回，单智能体任务成本降低 68%。xAI 公布 Colossus 1 拥有 15 万张 H100，Colossus 2 已有 11 万张 GB200，下周将再增 22 万张 GB300。
DAILY AI BRIEF 🗞 — Sept 25
OPENAI 🔥: - Upcoming ChatGPT Pro Max is $500/mo ($600 with VAT). Copy so far only adds Fastest Work and Codex; limits may still rise. - Likely Cerebras-backed for the Ultra Fast path. DevDay is Sept 29.
GOOGLE 🔥: - Gemini 3.8 Live with Live Avatar is GA in Gemini Enterprise. Lip-sync video agents, 97 languages, async tool calls mid-conversation. - Project Suncatcher flies a TPU prototype on SpaceX Transporter-18 with Planet, testing whether Google chips hold up in orbit.
ANTHROPIC 🔥: - Resumed charging for safeguard blocks that fire before Claude replies: biology, distillation attacks, and frontier LLM work. - 99.7% of Claude Code, https://t.co/w5NpeYznzO, and Cowork accounts did not hit the new billable blocks. Classifiers tuned under 0.1% false positives.
PERPLEXITY 🔥: - Fast Search is live in the Search API on Photon, a new Rust retrieval stack. 95% of results land in 230 ms or less. - 68% lower cost per agent task vs the default preset. Photon p99 dropped from ~800 ms to ~65 ms. - Portable Computer is in the Windows app for Consumer and Enterprise Pro and Max. Local harness; cloud model only with permission.
SPACEXAI 🔥: - Elon: Colossus 1 is 150k H100, 50k H200, and 30k GB200. Colossus 2 is 110k GB200 and 440k GB300. - Another 220k GB300 next week, 220k in November, and maybe 220k more by late December.
MICROSOFT 🔥: - Foundry adds model-agnostic voice agents on the same enterprise systems, tools, and controls stack.
META 🔥: - Muse Code usage limits reset for Connect. Model API and Muse Code hit Oracle Marketplace.
* Used Grok to compose this brief, cherry-picking the news and doing some post-editing. ** This daily brief also arrives in the daily email format; subscribe on the blog.
