Changelog — November 10, 2026

Claude Fable 5.2 (claude-fable-5-2) launched on 2026-11-09 at $8 input and $40 output per million tokens, down from $10 and $50 for Fable 5.1. It keeps the 1M-token context window and 128K maximum output, runs with always-on adaptive thinking and defaults to high effort. Cache reads cost 2.5% of the input price ($0.20 per million). Anthropic's models overview now lists Fable 5.2 as the Fable model, moves 5.1 to its legacy list, and gives 5.2 a retirement date of no sooner than 2027-11-09. It is on the Claude API, Claude Platform on AWS and Claude Models on Google Cloud (Anthropic managed), and also on Amazon Bedrock, Google-managed Claude Models on Google Cloud and Microsoft Foundry. Like Fable 5.1 it requires 30-day data retention and is not available under zero data retention unless Anthropic authorises it.

Fast mode and a removal date. Fast mode for Fable 5.2 delivers up to 2.5x higher output tokens per second at $16 input and $80 output per million tokens; set speed: "fast" on the request, with no beta header. Fast mode for Claude Opus 4.8 was deprecated on 2026-10-30 and is removed on 2026-11-30: afterwards speed: "fast" requests to claude-opus-4-8 return an error rather than falling back, and Anthropic points to Opus 5.5 as the migration. Claude Mythos 5.2 (claude-mythos-5-2) is limited to organisations verified through Anthropic's verification programs.

Other API changes to check. A request that edits earlier content and sends back a thinking block now returns a 400 for every account (the prefix check). A request that declares the earlier computer_20251124 tool returns a 400 on every platform except Amazon Bedrock. On the advisor tool (beta), a Fable 5.2 or Mythos 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor. Also in the window: an IP allowlist for organisations in the Claude Console (beta, 2026-11-05), an Admin API call that signs a Claude Enterprise member out everywhere (beta), and a beta for smaller requests in long Messages API conversations (message-threads-2026-08-12 header; a thread expires 5 minutes after its last request, 2026-11-02). On the Claude apps side, Personal Finance in Claude arrived in beta on 2026-11-05 for Pro, Max and Ultra plans in the US only, alongside new Enterprise security and admin controls, and project spend tags followed on 2026-11-06.

Models overview → Affects: /claude-cowork/pricing/, /claude-cowork/vs-api/, /tools/model-pricing/, /tools/cost-calculator/

Gemini 4 Carbon (gemini-4-carbon) became generally available on 2026-11-10 with the same introductory pricing as Gemini 4 Argon: $2.00 input and $10.00 output per million tokens (output includes thinking tokens) through 2026-11-30, then $4.00 and $20.00 from 2026-12-01. Google's pricing page lists no separate long-context tier for either model. Gemini Omni 1.2 Flash (gemini-omni-1.2-flash) went GA on 2026-11-05 with video generation of up to 15 seconds and conversational editing, at $1.50 input and $9.00 text or $12.00 video output per million tokens; requests to gemini-omni-1.1-flash route to it from 2026-12-03.

Deprecations to plan for. On 2026-11-02 Google deprecated gemini-2.5-pro and gemini-2.5-flash: neither is selectable in Google AI Studio, and requests route to gemini-3.8-flash and gemini-3.5-flash-lite respectively from 2026-12-01. gemini-embedding-2-preview routes to gemini-embedding-2 at the same rates from 2026-11-06. Gemini 3.8 Flash's introductory price ($0.75 / $3.75) runs through 2026-12-31 and doubles from 2027-01-01. The Computer Use tool reached GA on 2026-10-30.

OpenAI. GPT-6-Cyber (gpt-6-cyber, 2026-11-05) is for finding vulnerabilities and verifying fixes in authorised test environments and needs the Daybreak setup before a first request. GPT-Image-2.6 Sunburst and Flare (2026-11-04) add adjustable complexity, SVG output and animated GIF generation. GPT-6 Rosalind (2026-10-30) is for life-sciences research and limited to organisations in the trusted-access program. OpenAI listed no price changes in this window.

Gemini pricing → Affects: /tools/model-pricing/, /tools/cost-calculator/, /chatgpt/pricing/

2.1.319 adds permissions.denyInsteadOfPrompting: a tool call that would wait on a permission prompt is denied and the turn continues, in terminal sessions, subagents and teammates. It also adds /claude-api latency-optimize (finds where your Claude API code makes people wait and tries faster settings against your eval), support for MCP tasks in the terminal (a long-running MCP tool call moves to the background and Claude is told the result when it finishes; CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND makes it wait instead) and CLAUDE_CODE_DISABLE_FEATURE_FLAGS, which stops feature-flag fetching without turning off telemetry. 2.1.321 adds Claude Fable 5.2 as the default Fable model and moves saved Fable 5.1 picks to the fable alias (at once on Pro and Max, on Team and Enterprise once the organisation allows it; /model claude-fable-5-1 switches back). 2.1.321 and 2.1.322 add claude remote-control --settings <file> and fast mode on Vertex in the global, US and EU regions.

Behaviour changes to check. Self-paced /loop now ends when a turn finishes without rescheduling instead of scheduling a fallback wake-up (CLAUDE_CODE_LOOP_KEEPALIVE=1 keeps the old behaviour). Behind an LLM gateway or with telemetry or feature flags off, auto mode now asks about or denies calls the server cannot check instead of falling back to the built-in classifier (CLAUDE_CODE_AUTO_MODE_SERVER=0 keeps the classifier). 2.1.317 added a check that asks before a web or shell tool call sends your account email address to the network, and 2.1.321 narrowed it to refuse only a shell command that holds your email and names a User-Agent, so scheduled runs no longer stop at its prompt. 2.1.317 removed the Read tool's "file unchanged since your last Read" stub, so reading an unchanged file returns its content again. 2.1.318 moves a background subagent's shell command to the background after 3 minutes, stops installing the IDE extension when started in an IDE terminal, and makes deniedMcpServers URL entries also block matching Anthropic-provided servers. The built-in Claude in Chrome tools lost shortcuts_list and shortcuts_execute in 2.1.319.

Smaller additions. 2.1.316 adds CLAUDE_CODE_SUBAGENT_EFFORT (the effort subagents run at when nothing else names one) and CLAUDE_CODE_DISABLE_ULTRATHINK, and deprecates fast mode for Opus 4.8 (see the Anthropic entry). 2.1.317 makes subagents on a different model use that model's saved effort level. 2.1.318 adds the date a stored claude.ai sign-in ends to claude auth status (loginExpiresAt) and a stderr warning from claude -p ten days before it ends. The Claude apps gateway gained pooled group budgets, a layered policy mode and a cost report endpoint across these releases.

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

2026.11.2 (2026-11-08) is a patch release with no new features, but its notes list required actions. After a manual source upgrade, run openclaw doctor --fix before startup so older scheduled jobs get a stored inference account. Existing-session Chrome MCP profiles now reject HTTP and HTTPS navigation and tab opens; use an OpenClaw driver profile for strict navigation. For affected browser extension profiles, move the extension relay secret file aside and run openclaw browser extension pair to re-pair, and note that extension setup now rejects group-writable package files from shell installers. The approvals commands now refuse to run when both --gateway and a node target are given. Blank cron edit flags now clear old failure-alert channel, destination or account overrides. Sandboxed web search enforces the provider trust policy, so if an excluded third-party provider was selected, enable a bundled or verified official one. Global installation with Bun 1.4.4 now works when Node is not installed. Plugins that declare the updated API floor need 2026.11.2 or newer, and affected 2026.10.4 installs need the update recovery guidance.

2026.11.1 (2026-11-05) stops failed Mac settings saves from overwriting secrets and keeps Gateway authentication intact when Tailscale exposure is turned off. Saved subscription accounts can make configured Claude CLI models available without a separate native login, and Anthropic API-key setup can record an optional workspace ID. Gemini Flash text models use the Interactions API by default on the official API-key endpoint (choose google-generative-ai when explicit Gemini cache handles are required). It also hardens network paths: ClickClack and ClawHub reject redirects into unrelated private services. 2026.10.10 shipped on 2026-11-03; its notes are published as a link to the docs release page and we have not itemised it here.

Not yet counted. A v2026.11.3 tag appeared in the release feed on 2026-11-10 but has no GitHub release or docs release page yet, so we track 2026.11.2 as current. The project also published two posts: "Use new models in OpenClaw without an update" (2026-11-06), saying OpenClaw learns about new models from a hosted catalogue, your provider account and an installed Codex so a model launch no longer waits for a release, and "OpenClaw Multiplayer" (2026-11-05). The excerpts name no setting or command for either, so check the posts before relying on them.

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

v0.23.0 ("The Hearth Release", 2026-11-05) serves your Desktop workspace in a browser tab from your own gateway. The Webapp is off by default; enable it with webapp.enabled or Web Access. Each machine now runs one gateway that owns every session, so the CLI, TUI, Desktop, editors, bots and cron stop running separate copies of the agent against the same chat. New: /rewind (or Esc twice) to edit and resend an earlier message or fork from it; hermes plugins update and an opt-in plugin auto-update switch; price tags and spend limits on paid MCP tools; hermes curator lock to stop the agent rewriting a skill you curated; /model reset; hermes sessions refresh-prompts; and an opt-in compression.prepare_ahead: true. The silent default model moves from GLM-5.2 to GLM-5.3, and the built-in skill set shrinks from 58 to 44 with Word, PowerPoint and Excel merged into one microsoft-office skill.

Breaking changes to check. The old built-in browser tools are gone: browser automation runs on browser_exec only, and agent-browser and the ten browser_* tools (browser_navigate, browser_click and the rest) were removed, so prompts, skills or toolset lists naming them need updating (homes that used Camofox get the camofox plugin installed automatically). hermes mcp serve is now hermes mcp-serve, and MCP clients that launch Hermes as a server need their command changed by hand. Desktop needs macOS 13+ and 64-bit Windows after the move to Electron 44.

Patch releases. v0.22.6 (2026-11-02) and v0.22.7 (2026-11-03) are roll-ups with curated notes deferred to 0.23.0. v0.23.1 (notes dated 2026-11-09, tagged 2026-11-10) rolls up about 618 merged PRs, and fixes a plugin-catalog problem: the published plugin-catalog.json had grown past the 2 MiB cap clients accept, so every install silently fell back to the bundled catalog and hermes plugins install <name> could not find newer plugins. Update with hermes update or nousresearch/hermes-agent:stable. Full curated notes for the window ship with v0.24.0.

Releases → Affects: /hermes/, /hermes/setup/, /hermes/mcp-tools/, /hermes/memory/, /hermes/skills-guide/

v0.0.141 (2026-11-03) sends operation telemetry to NVIDIA's Production receiver by default. Eligible installer and supported CLI operations attempt to send completed-operation telemetry; v0.0.140 had pointed the same default at a fixed TEST receiver. Set NEMOCLAW_DISABLE_TELEMETRY=1 in each shell before installation or CLI use to disable collection and delivery. NVIDIA's best-practices page lists the collected fields and exclusions. The release also lets managed Hermes onboarding accept --switchyard-config <file> to generate routing configuration for efficient and capable hosted models (credential attachment for the routed targets is still tracked in issue #13863, so it does not establish production routing), and makes custom inference endpoint setup refuse conflicting network policy rules before changing the provider or route.

v0.0.142 (2026-11-06) fixes native inference onboarding. New managed single-host vLLM containers publish their host port only on loopback and the private OpenShell Docker bridge gateway; existing containers keep their bindings until recreated, and managed route recovery rejects owned bearerless containers with unsafe bindings. OpenClaw Model Router onboarding attaches a sandbox-scoped native provider, pinned NVIDIA-compatible provider profiles are supported without keeping the bundled NVIDIA network grant after attachment, and host assessment now finds the WSL-provided nvidia-smi when it is absent from PATH. In the repository, a commit on 2026-11-10 retires the deprecated Model Router; it is not part of a release yet.

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

Kilo Code 7.8.27 is the new stable release (2026-11-05), nine stable-line versions after 7.8.18. Its GitHub release notes read only "No notable changes", so there is nothing to act on beyond updating. Pre-release tags 7.8.26 (2026-11-04) and 7.8.28 (2026-11-07) were also published; we track stable releases and record 7.8.28 as the pre-release ahead of stable.

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