OpenAI and Anthropic both bet on cheaper, longer-lived context

Cheaper context, longer cache

Both labs spent the week attacking the cost of keeping context around. OpenAI's API now offers 12-hour prompt caching, so shared instructions, tools, and context stay warm even when requests are hours apart — cache writes at that duration cost 2.5x the input token price, a trade worth running the numbers on. The Decisions API went generally available on gpt-6-luna and took direct advantage of it: shared input can be cached across prompt-cache calls, question instructions can carry images, structured data flows in through input_json, and a single request now handles up to 400 questions instead of 200. OpenAI framed the offering simply as fast, focused, low-cost decisions for any application.

The SDKs were reworked around those Decisions types, and the changes are breaking in minor releases. The Go SDK moved Decision and Live instructions to typed unions, so plain string assignments stop compiling. The Node SDK made question instructions a string or parts array, which means question.instructions.trim() can now blow up on the array form, and code extending DecisionInputMessage has to switch to the namespaced variant. Python took the same structured instructions and richer input message roles, with string handling preserved but now needing an isinstance check. Vertex AI support for the Responses API landed across Node, Python, and Go — and separately as a marked Go Vertex client milestone — alongside voice consent management, voice deletion, and 30-minute cache-write usage counts. Long tail: Node's WebSocket truncation fixes, Go's X.509 and WebSocket restore fixes, voice consent in Go, raw WebSocket helpers, and per-attempt auth timeouts.

Claude Code's security week

Claude Code shipped a fistful of updates, and the throughline was sandbox and permission hardening. Sandboxed Bash commands on macOS and Linux no longer run shell startup files outside the sandbox, symlinked denyWrite globs now apply at real paths, and sandboxed commands can't touch the MCP config passed via --mcp-config. Linux users got nested user namespaces as an opt-in — with the honest caveat that it lets them get around the sandbox's file protections — and a fix for a regression where Vertex and Bedrock requests failed on tool search rejection. Elsewhere: worktree removal no longer drops commits when HEAD is detached, plus /auto-mode-setup; a new OpenTelemetry safety event and /claude-api api-version for API version 2026-10-21; cloud session login and prompt-too-long fixes; background cleanup patches plus a Monitor tool on Bedrock, Vertex, and Foundry; and a round of reliability fixes.

Anthropic's breaking API version, and an open standard

Anthropic released API version 2026-10-21 with a migration guide, and every official client now defaults to it: Python and Ruby hit 2.0.0, TypeScript and PHP 1.0.0, Go 2.0.0, Java 3.0.0, C# 13.0.0, and the ant CLI 2.0.0. The SDKs carry breaking changes independent of the API version and are now Apache 2.0 rather than MIT — worth a license review before upgrading. On the platform side, Claude Models on Google Cloud lets teams reach the Anthropic-operated API through a GCP project with IAM and Marketplace billing, with Fast mode for Opus 5.5; organization web fetch controls add domain allow and block lists for admins; and inference hooks graduated from beta alongside /v1/user_profiles and org webhooks. Governance news: Anthropic is donating Agent Skills to the Agentic AI Foundation, now integrated by more than 70 agent platforms. And Claude Duet, a proactive-agent app in research preview for Pro and Max plans, is the week's most consumer-facing bet, padded out by two routine app maintenance releases.

Workspace controls across ChatGPT plans

ChatGPT's business tiers converged on the same features: locked chats on Business and locked chats on Enterprise and Edu hide conversations behind a PIN or device auth without breaking retention policies, Enterprise gained Record for meetings and voice notes and shareable workspace profiles, and Business admins got member management in Admin Console. Enterprise analytics were reorganized into Usage and Insights, with a 7-day migration grace period for users leaving personal workspaces. Meanwhile ChatGPT Ads expanded to 26 more countries.