Agent Delta #1: Claude Threads, Gemini 2.5 sunset, Computer Use goes GA

TL;DR: Long-session latency just got cheaper to fix, Gemini 2.5 model strings stop meaning what you think on 1 Dec, and computer-use agents get a GA tool. Action this week: grep your code for gemini-2.5-pro, gemini-2.5-flash and Opus 4.8 fast mode.

1. Anthropic Message Threads (public beta, 2 Nov)

What shipped: add a thread field and the message-threads-2026-08-12 beta header; send the full conversation once, then only new messages plus the previous response ID. Anthropic reports about 35% lower median time-to-first-token on 500k-1M token requests. Token billing unchanged.

Why you care: long-running coding and computer-use loops get faster without a rewrite. Catch: threads expire 5 minutes after the last request and are not ZDR eligible.

Do this: use the SDK helper client.beta.messages.thread(...), which replays the full conversation for you when a continue request returns 404. Source · Release notes

2. Gemini 2.5 Pro and Flash deprecated, auto-rerouted 1 Dec (2 Nov)

What shipped: gemini-2.5-pro and gemini-2.5-flash are gone from AI Studio. From 1 Dec, API calls to them silently route to gemini-3.8-flash and gemini-3.5-flash-lite, billed at those rates.

Why you care: Google lists breaking changes on the new models: function-calling and thought_signature validation, thinking_level replacing thinking_budget, removed sampling params and prefill. Agents with tool calls can break without a code change.

Do this: pin the new model string in a branch now and rerun your tool-calling evals before December. Source

3. Gemini Computer Use is generally available (30 Oct)

What shipped: Google's Computer Use tool reached GA across the Interactions and GenerateContent APIs, for browser, mobile and desktop control agents.

Why you care: a GA, supported screen-control tool from a frontier lab is a real option for production computer-use agents. You still run the execution environment client-side.

Do this: run your ten hardest browser tasks before switching; keep your current model as fallback. Source

4. Claude Code 2.1.316 (31 Oct) and the Opus 4.8 fast-mode deadline

What shipped: CLAUDE_CODE_SUBAGENT_EFFORT sets the effort level for subagents instead of inheriting the session's; CLAUDE_CODE_DISABLE_ULTRATHINK turns off the ultrathink keyword; pooled group budgets and layered policy mode for the Claude apps gateway. Opus 4.8 fast mode is going away; the notes say to migrate to Opus 5.5 to keep the speedup.

Why you care: per-subagent effort is a direct cost lever for multi-agent runs, and gateway budgets help teams cap spend.

Do this: set subagent effort low for search and lint subagents, then compare cost per task. Source

5. Google Cloud Developer plugin lands in Antigravity Marketplace (2 Nov)

What shipped: agy plugin install google-cloud-developer adds Google Cloud onboarding, auth, safe gcloud guidance and a pre-authorized Developer Knowledge MCP server for live official docs.

Why you care: grounding a coding agent in current cloud docs, plus guardrails on destructive CLI commands, is exactly the plumbing most teams hand-roll.

Do this: install it in a staging project and test a deploy flow end to end before giving it real IAM. Source

Quick hits

• Honeycomb: Personal and Team AI SRE agents in early access, with MCP connectors (3 Nov).

• Snowflake: always-on CoCo and CoWork Bots announced at Expedition 2026 (3 Nov).

• Oumi: agent build-and-train platform with Bring Your Own Harness, DPO and GRPO (4 Nov).

• Cursor Origin API: repository.renamed webhook and app-declared IP allowlisting (4 Nov).

Agent Delta: what changed in AI this week, and what to change in your agent.