Thursday roundup: 22 Oct 2026

Anthropic releases Claude API version 2026-10-21 with new major SDKs, Claude Code 2.1.307 adds a safety event and an MCP retry fix, OpenAI models arrive on Google Cloud in private preview, and OpenAI adds per-project service tier controls.

  1. Anthropic

    Claude API version 2026-10-21 and new major SDKs

    Released 21 Oct 2026

    Each new SDK uses the new API version by default: Python 2.0.0, TypeScript 1.0.0, Go 2.0.0, Java 3.0.0, C# 13.0.0, PHP 1.0.0, Ruby 2.0.0, and the ant command-line tool 2.0.0.

    The SDKs also include breaking changes of their own, and the licence has changed from MIT to Apache 2.0.

    Separately, user profiles now have enforcement_status and enforcement_actions fields, which show when Anthropic has restricted or banned one of your end customers.

    Who it affects Anyone with unpinned Claude SDK versions, since a routine dependency update will pull in the breaking changes. Platforms and gateways can now handle one customer's ban without the whole organisation being affected.

    Source: Claude Platform release notes ↗

  2. Anthropic

    Claude Code 2.1.307

    Released 21 Oct 2026, 23:17 UK

    /claude-api api-version updates your Claude API code to the new 2026-10-21 version.

    A new OpenTelemetry event, claude_code.safety_event, is recorded whenever a permission rule, auto mode, a safety check, a hook or the sandbox blocks, asks about or sends back a tool call.

    If a connection drops partway through a call to an MCP server over HTTP, the call is no longer sent twice. Only read-only or idempotent tools are retried now.

    Checkout hooks on self-hosted runners now time out after 30 minutes by default, which you can change with --checkout-hook-timeout-min. @claude review stop cancels an always-on Code Review subscription.

    Versions 2.1.306 and 2.1.308 were bug fixes only.

    Who it affects Teams that want an audit trail of when agent guardrails stepped in, and anyone using HTTP MCP tools that change things, which no longer risk running twice.

    Source: Claude Code 2.1.307 release notes ↗

  3. OpenAI and Google

    OpenAI models on Google Cloud

    Released 21 Oct 2026

    GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are rolling out as a private preview in Model Garden, Google Cloud's model catalogue. General availability is planned for the coming weeks.

    They are bought through Google Cloud Marketplace, and the spend counts towards any existing Google Cloud commitment. OpenAI says its models are coming "soon" to Antigravity and Gemini Agent, but that has not shipped yet.

    Who it affects Organisations on Google Cloud, which can now use GPT, Gemini and Claude on one cloud with one bill and one set of governance controls.

    Source: OpenAI announcement ↗

  4. OpenAI

    Service tier controls for each project

    Released 21 Oct 2026

    Admins can now choose which service tiers, such as Ultrafast, each project is allowed to use. They can also restrict tiers for the whole organisation in Platform settings.

    Who it affects API admins who want to stop one project from quietly running everything on the more expensive tiers.

    Source: OpenAI API changelog ↗

← All news