Week in review: OpenAI ships managed Agents API, Apple's new Siri reportedly runs on Gemini, and three vendors add agent spend controls

Consolidating what actually shipped this week, since it feels like more of a pattern than usual. OpenAI put its Agents API into public beta — managed orchestration, long-running sessions, context management, with sandbox compute available from OpenAI or partners (Vercel, DigitalOcean mentioned). They also started testing sponsored agents inside ChatGPT with Wayfair and Angi as the first advertisers, and added a ChatGPT sidebar integration for Microsoft Word. GPT-5.5 is scheduled for retirement on October 14 across ChatGPT, ChatGPT Work and Codex. Apple released Siri AI — personal context, onscreen awareness, systemwide app actions, standalone app. English beta now, five more languages in October. Multiple outlets report it's running on Google's Gemini under the hood. Google shipped Gemini 3.8 Flash (third Flash release in six weeks) plus a security-focused model for government/enterprise. Gemini Enterprise pricing changed: pay-as-you-go, up to 20% token discounts, monthly caps on agent spend, and a $0 base subscription tier. xAI launched Grok Bot — persistent agents with memory, dedicated environments, browser access, and the ability to coordinate with other bots. Enterprise version includes access, network and audit controls. Microsoft expanded Copilot governance (Outlook, Purview, SharePoint) and added Power BI / CSV / TSV as Copilot Notebooks knowledge sources. Anthropic claims Claude now leads 26% of its internal R&D, up from roughly zero at the start of the year (Bloomberg, Sept 17). Open source: Nvidia reportedly agreed to acquire Hugging Face (~$13B, announced late August). Separately there's a reported DNS rebinding attack that can persistently poison a local Ollama agent via the browser — worth reading if you run local inference with tool access. My take as someone building on top of these APIs: The Agents API is the one that changes my roadmap. A meaningful chunk of what I've written is session persistence, context management and sandboxing — all of which is now a managed service. I don't think that's a disaster, but it does mean nobody should be pitching orchestration as differentiation anymore. The more interesting signal is vocabulary. Audit controls, spend caps, admin governance — all in the same week, from three different vendors. That's what a category looks like when procurement gets involved. If you're selling agent tooling and you don't have audit logs, that's your next quarter. On Apple/Gemini: if it holds up, it's the strongest evidence yet that frontier model training is a three-or-four-company game and everyone else is a customer. That's not necessarily bad — it's how databases went — but it should kill the "we'll train our own foundation model" line in a lot of decks. One caveat on the Anthropic 26% figure: self-reported, from a vendor, with no methodology published. Directionally interesting, not evidence. Curious whether anyone here has actually run the Agents API beta yet — specifically how the sandbox partner setup behaves for long tasks, and whether the context management is genuinely better than rolling your own.