Changelog — October 26, 2026

Claude Duet is a research preview of a proactive agent (2026-10-21). It is a separate app on iOS, Android and claude.ai, built around Claude Opus 5.5, available by waitlist to US Pro and Max subscribers and included in those plans. It runs on its own virtual machine with a browser, sends a morning brief, checks your connected accounts every few hours, and shows you a message, payment, signature or deletion and waits for approval first. Some actions are off-limits even with approval, such as trading securities or moving money between accounts. It uses your normal Claude usage limits, and background checks count against them. Anthropic lists over 6,000 connectors and plugins from the Claude Marketplace, and says it cannot yet show its browser and that personal errands are less developed.

API version 2026-10-21 ships with new SDK majors. The default-version SDKs are Python 2.0.0, TypeScript 1.0.0, C# 13.0.0, Go 2.0.0, Java 3.0.0, PHP 1.0.0, Ruby 2.0.0 and the ant CLI 2.0.0. They carry breaking changes that don't depend on the API version and are now Apache 2.0 licensed instead of MIT. Claude Code 2.1.307 adds /claude-api api-version to move your code to the new version. User profiles gain enforcement_status and enforcement_actions fields, and an enforcement_status[] list filter, to show whether Anthropic has restricted or banned an end customer.

Google Cloud (2026-10-22). "Claude Models on Google Cloud (Anthropic managed)" is the Claude API operated by Anthropic and reached through your Google Cloud project, with Google Cloud IAM for access and Marketplace billing. Fast mode is available for Claude Opus 5.5 on it. The API notes announce no deprecations.

API release notes → Affects: /claude-cowork/, /claude-cowork/pricing/, /claude-cowork/vs-api/

2.1.309 is the release to read if you rely on the sandbox or permission rules. It fixes sandboxed Bash commands on macOS and Linux running the shell's startup files outside the sandbox before each command, macOS denyWrite glob patterns not applying through a symlink, and sandboxed commands being able to modify the MCP config file passed with --mcp-config. It also closes several permission-check gaps (an Edit(path) deny rule not blocking a Bash redirect in some commands, compound cd and git commands not prompting) and makes sed -i use the standard Bash prompt, which means /rewind no longer restores it.

Behaviour changes to check. claude ultrareview now requires workspace trust: it asks on a terminal and refuses otherwise, so scripts must pass --trust-workspace (no value) or set CLAUDE_CODE_DISABLE_ULTRAREVIEW_TRUST_CHECK=1 (2.1.309). In bypass permissions mode in the Desktop app and VS Code, the dangerous rm and rmdir prompt now denies at once with a rewrite hint. Self-hosted runner checkout hooks stop after 30 minutes by default; raise --checkout-hook-timeout-min or set it to 0 (2.1.307). The /claude-api skill now builds Managed Agents subagent teams on multiagent_20261001 with dynamic workflows on by default.

Additions. A claude_code.safety_event OpenTelemetry event was emitted (2.1.312 stopped sending it again, so do not build on it) for each tool call that a permission rule, auto mode, a safety check, a hook or the sandbox stopped, prompted or sent back (2.1.307). The Monitor tool and plugin monitors now work on Bedrock, Vertex and Foundry (2.1.310). @claude review stop (also off and pause) ends an @claude review always subscription on a pull request (2.1.307). 2.1.308 lists only bug fixes and reliability improvements.

Changelog → Affects: /claude-cowork/setup/, /claude-cowork/faq/claude-code-undocumented-configuration/, /commands/

2026.10.7 is a patch release (61 PRs, 54 contributors) that fixes updates, recovery and lost replies. A suspiciously truncated openclaw.json no longer overwrites the healthy recovery copy, so Doctor can restore your settings. Package updates stop before staging a release that needs a newer Node than the Gateway service uses, and the Gateway keeps running. Restoring a backup keeps supported older exec approval rules instead of replacing them with deny-all, and malformed rules stay denied. Replies queued before a plugin reload can use the current provider, and agents can continue unfinished tool work after an empty, length-limited response.

A migration note for Claude CLI users. Managed Claude CLI turns now use OpenClaw's tool surface. After upgrading they stop loading Claude user settings, CLAUDE.md and rules, auto-memory, user hooks and plugins, Chrome, and Claude-only MCP servers; OpenClaw-configured MCP servers remain. Existing CLI conversations start one fresh native Claude session on their first turn, and a change in available tools does the same. The notes also tell you to check which Node executable the Gateway service uses, since it may differ from your shell. A 2026.9.34 tag appeared the same day, but the 2026.10.7 notes do not mention it.

Release notes → Affects: /openclaw/, /openclaw/setup/, /openclaw/troubleshooting/, /openclaw/configuration/

0.22.4 (2026-10-21) is a small fix release; 0.22.5 (2026-10-22) rolls up about 81 PRs. 0.22.4 removes the suggestion pills above the Desktop composer and skips the dependency consent step for plugins that install nothing. The 0.22.5 notes deliberately list items without detailing them: a reworked Nous sign-in, hermes update holding back a stable release it refuses instead of offering it, pluggable delegation backends (route subagents by model map or profile, switchable per session), plugin slash commands that can ask you to pick an option, hermes import-agent and an OpenClaw migration, and hermes mcp serve and the Singularity backend moving out of core into plugins. Full curated notes arrive with v0.23.0, so treat the list as a lead.

Update with hermes update or the installer one-liner; Docker and Hermes Cloud use nousresearch/hermes-agent:stable or :latest, and the Microsoft Store app updates through the Store. The installer-based desktop app does not change until the next bundled release. The plugin catalog also gained bubblewrap (a bwrap terminal backend on Linux), Singularity/Apptainer, mcp-serve and agent-importers.

Releases → Affects: /hermes/, /hermes/setup/, /hermes/mcp-tools/, /hermes/dashboard/

7.8.18 changes how remote MCP servers connect. After the MCP TypeScript SDK update to 1.32.1, remote MCP servers that redirect to another origin, or that answer POST with a 301, 302 or 303, no longer connect; use the server's final URL instead. It also exposes hosted cloud sessions through the local Kilo API so editors can find them without handling provider credentials, and fixes copying chat text in VS Code 1.143. 7.8.16 (a pre-release) restricts browser links and MCP OAuth authorization URLs to HTTP(S), blocks compromised eMudhra certificate chains via a pinned Bun runtime, and treats shell expressions in MCP prompts as literal text. 7.8.17 lists no notable changes.

Releases → Affects: /kilocode/, /kilocode/setup/, /kilocode/security/

v0.0.138 tightens installation and image-build security. Linux system-mode Ollama installation verifies a pinned official installer before running it and stops if the download can't be verified, and the repository .dockerignore now excludes credential, key and certificate names at every directory depth. New local-inference profiles for host.openshell.internal and host.docker.internal admit IPv6 unique-local addresses as well as private IPv4; existing sandboxes keep their recorded permissions until you reselect the local provider. NEMOCLAW_STATUS_PROBE_TIMEOUT_MS and NEMOCLAW_SANDBOX_EXEC_TIMEOUT_MS values above 2147483647 now fall back to their defaults instead of timing out immediately, and closing stdin during onboarding cancels with a nonzero exit while preserving the session for nemoclaw onboard --resume.

Release notes → Affects: /nemoclaw/, /nemoclaw/setup/, /nemoclaw/local-gpu/

OpenAI's Decisions API is generally available with gpt-6-luna (2026-10-22). It adds prompt-cache discounts for shared input, image support in question instructions, and a limit of up to 400 questions per request; the changelog lists no prices. On 2026-10-21 OpenAI also let projects choose which service tiers they can use in Platform settings, with organization settings able to restrict tiers for every project.

Gemini storage default (2026-10-26). New paid-tier projects using the Generate Content API (generateContent, streamGenerateContent, batchGenerateContent) now store requests by default (store=true), matching the Interactions API; existing projects keep their current setting. If you send sensitive prompts from a new project, set store explicitly. Gemini 4 Argon's price is unchanged ($2 input and $10 output per MTok through 2026-11-30, then $4 and $20), and we re-checked the Google rates in our tables on 2026-10-22.

New on OpenRouter, not yet confirmed by the vendors. moonshotai/kimi-k3.1-pro (listed 2026-10-23) and minimax/minimax-m3.1-flash (2026-10-19), both with a 1,048,576-token window, appeared in the catalogue. Moonshot's own documentation did not mention K3.1 Pro on 2026-10-26, so we have not added either to the price tables. Our video summary of a five-build K3.1 Pro versus Opus 5.5 test (in our news section) has one creator's results. xAI's Grok Bot docs overview was updated on 2026-10-22; it lists Tag @bot on X, group chats and Team Bots and names no new release.

OpenAI changelog → Affects: /chatgpt/pricing/, /grok-bot/, /tools/model-pricing/