Claude API message threads in beta, Gemini 2.5 Pro and Flash auto-reroute on Dec 1, Vercel agent-desktop 1.0

Message threads let the Messages API hold a conversation on the server for a short while. Your first request looks the same as today, with the full conversation, system prompt and tools, plus a thread parameter set to create. After that you send only the new messages and the previous response id with continue, and the API rebuilds the conversation before calling the model. A thread expires 5 minutes after its last request and is deleted within an hour. You need the message-threads-2026-08-12 beta header, though the SDK's thread helper adds it and handles replays for you.

thread = client.beta.messages.thread(model="claude-opus-5-5", max_tokens=1024)
thread.send({"role": "user", "content": "Hi!"})

In a long agent loop, every turn re-uploads the whole conversation, and the API validates and tokenizes all of it again. Anthropic says threads cut that work down to the new turn, so time to first token stays lower and steadier as the conversation grows. Threads are best-effort, though: a continue call can return a 404 at any time. Your app still has to keep its own copy of the conversation and a path to resend it in full.

Google has deprecated gemini-2.5-pro and gemini-2.5-flash and removed them from the Google AI Studio model picker. Starting December 1, gemini-2.5-pro calls will route to gemini-3.8-flash, gemini-2.5-flash calls will route to gemini-3.5-flash-lite, and both will be billed at the new model's rates. API access to gemini-2.5-pro is now limited to existing users who have actually been using it.

Your requests keep succeeding even if you change nothing, but a different model answers them, so quality, latency and cost can shift without warning. Google lists the breaking changes: function calling and thought_signature validation, thinking_level replacing thinking_budget, and the removal of sampling parameters and prefilled model turns. If you have prompts or eval sets tuned for 2.5, rerun them against the new models in November.

agent-desktop is a CLI and C library that lets AI agents read and operate desktop apps on macOS, Windows and Linux through the OS accessibility tree, the structured list of on-screen elements the operating system exposes. A snapshot returns JSON with a ref for every element an agent can act on. The agent clicks a ref, then takes another snapshot to check the result. No pixel matching is involved.

npm install -g agent-desktop
agent-desktop snapshot --app Finder -i

It's lighter than screenshot-and-click computer use, and the JSON output is easy to wire into coding agents like Claude Code or Codex as a tool. The loop stays inside your own agent, so you aren't tied to one model. Snapshots and clipboard reads can contain private screen content, so follow the README and keep them out of logs.