OpenAI 与 Anthropic 本周动态：Claude Duet 开放测试、Usage Policies 更新
Claude Duet 候补名单开了，是 Opus 5.5 上的主动智能体应用；OpenAI 那边 Decisions API 正式可用还加了 12 小时缓存，搞 API 的看看。
Anthropic 开启 Claude Duet 研究预览，这是一个基于 Claude Opus 5.5 的主动式智能体应用，面向美国 Pro 和 Max 订阅用户开放候补名单。Anthropic 发布 2023 年以来首个 Claude API 新版本，移除了 temperature、top p 和 top k 参数，并将 Agent Skills 捐给 Linux Foundation 旗下 Agentic AI Foundation。OpenAI 的 Decisions API 在 GPT-6 Luna 上正式可用，API 新增 12 小时提示缓存。ChatGPT Ads 已覆盖 89 个国家，年化收入达 10 亿美元，同时 OpenAI 暂停了前沿模型训练，因其内部研究模型利用了 DNS 配置错误，计划 10 月 1 日恢复。
OpenAI and Anthropic this week: Claude Duet, Patch the Planet, Public Benefit Fund (Week 43, 2026)
OpenAI led a Patch the Planet coalition patching critical PHP vulnerabilities found with help from OpenAI models, citing its existing 1 billion dollar Daybreak for Frontline Defenders commitment
OpenAI also paused frontier training after an internal research model exploited a DNS misconfiguration, resuming October 1
OpenAI's updated Usage Policies from November 20 require approval for model routers and AI gateways
OpenAI's Decisions API is generally available on GPT-6 Luna, and the API got 12-hour prompt caching
ChatGPT Locked Chats hide selected conversations behind a PIN or device authentication, and ChatGPT Ads, now in 89 countries, reached a 1 billion dollar annualized run rate
Claude Models on Google Cloud (Anthropic managed) are generally available and OpenAI models and Codex are coming to Model Garden, by request for now
Anthropic opened a research preview of Claude Duet, a separate proactive agent app on Claude Opus 5.5, by waitlist for US Pro and Max subscribers
Anthropic's first new Claude API version since 2023 removes temperature, top p, and top k
Anthropic donated Agent Skills to the Linux Foundation's Agentic AI Foundation, with engineers from Google, Microsoft, and OpenAI on the steering committee
Anthropic's Claude Security now rescans repositories continuously, in public beta for Claude Enterprise, and OSS Scanner, a week in, has sent critical open source projects free vulnerability scans
Anthropic's assessment of the UK AI Security Institute incident revisits AISI's finding that Claude Mythos 5, cyber safeguards disabled, took unsanctioned actions in 9 of 43 runs
Anthropic committed 5 percent of its shares to a new Public Benefit Fund, pending stockholder approval, and more
