Home › Changelog › 2026-11-21
Last updated: 2026-11-21
Changelog — November 21, 2026
Two model launches and a stricter OpenClaw. OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, Google made Gemini 4 Neon generally available, and xAI ended grok-build-0.1. OpenClaw 2026.11.6 checks shell-command approvals more consistently, so some commands that ran automatically will now prompt or be denied. Claude Code 2.1.327 to 2.1.330 adds sign-in for organizations managed in Anthropic's Enterprise Admin Console and a way to keep MCP tokens off the command line. Kilo Code's new stable release makes Codestral the default autocomplete model.
2026-11-18
Models
OpenAI · Google · xAI — GPT-6.2 Astra and Sol, Gemini 4 Neon, grok-build-0.1 end of life
OpenAI added GPT-6.2 Astra (gpt-6.2-astra) and GPT-6.2 Sol (gpt-6.2-sol) on 2026-11-17 on the Responses and Chat Completions APIs. OpenAI pitches Astra for demanding reasoning and coding and Sol for everyday coding and agent workflows. Per OpenAI's pricing page, Astra costs $8 input, $0.40 cached input and $40 output per million tokens ($16 / $0.80 / $60 past 272K input tokens), below GPT-6 Astra's $10 / $50; Sol costs $2 / $0.10 / $10 ($4 / $0.20 / $15 long context), the same as GPT-6.1 Sol. Tool calling needs the Responses API, which also changes how reasoning is kept when a conversation prefix is edited and how the top-level instructions field is handled. The same day OpenAI put two-hour and eight-hour background completion windows into public beta for GPT-6.2 Astra and Sol, GPT-6.1 Terra and GPT-6 Luna, at 30% and 40% below Standard rates, and said new models will not support Flex processing. "Pro" variants of both models appear in the OpenRouter catalogue but not on OpenAI's pricing page, so we do not list them.
Gemini 4 Neon (gemini-4-neon) became generally available on 2026-11-18 at $3 input and $15 output per million tokens (output includes thinking tokens; context caching $0.15), with no introductory end date. It has four thinking levels (low, medium, default high and xhigh) and the same unified 1M-token context window as the rest of Gemini 4. Batch and Flex are $1.50 / $7.50 and Priority is $5.40 / $27. Google also made Genie 4 (genie-4) generally available on the Interactions API on 2026-11-19. Gemini 4 Carbon and Argon still move from $2 / $10 to $4 / $20 on 2026-12-01. xAI: grok-build-0.1 reached end of life on 2026-11-19 and is no longer served on the API; xAI names no replacement. Grok 4.7 and 4.6 prices are unchanged at $2 / $6.
Gemini pricing →
Affects: /tools/model-pricing/, /tools/cost-calculator/, /chatgpt/pricing/, /grok-bot/pricing/
2026-11-20
Claude Code
2.1.327 → 2.1.330 — Admin Console sign-in, tokens off the command line, MCP hook blocking
2.1.330 (2026-11-20) adds sign-in to organizations managed in Anthropic's Enterprise Admin Console: after claude auth login, Claude Code applies the organization's settings and uses its cloud provider or gateway. It also adds a built-in "easel" mod that draws charts when asked and under MCP metrics results (turn automatic charts off in /config), CLAUDE_CODE_FULL_SIGN_IN_LINK=1 to print the full sign-in link, and --start-dir for the self-hosted runner. It fixes Claude Code freezing for several minutes right after compacting a very long conversation, and WebFetch freezing the session on some pages.
2.1.327 (2026-11-17) adds claude mcp add-json -, which reads the server's JSON from stdin so tokens stay off the command line (not available inside a Claude Code session), and fixes claude mcp add and remove at project scope deleting other top-level keys such as $schema from .mcp.json. It also fixes Remote Control in worktree mode deleting a running session's git-ignored files, such as .env, when the server stopped. 2.1.328 adds "onFailure": "block" for mcp_tool hooks, so a failing hook blocks the action as command and HTTP hooks already can, plus auto mode for Opus 4.6 and Sonnet 4.6 on Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway. 2.1.329 adds a trust-dialog warning naming where a folder's hooks are declared, lists this session's scheduled tasks, /loop wakeups and forks in /tasks, and fixes ~/.claude.json sometimes being reset to defaults during a stalled save.
Releases →
Affects: /claude-cowork/setup/, /claude-cowork/faq/claude-code-undocumented-configuration/, /commands/
2026.11.6 (2026-11-21; 386 pull requests, 157 contributors) is a patch release, but its security changes can change what runs unattended (release notes). Gateway shell commands that need approval are now checked more consistently on macOS and Linux, and commands that change their environment are checked too, so a command that previously ran automatically may now prompt or be denied under your existing policy. Reapplying an unchanged exec policy no longer silently disables automatic command review. OpenClaw no longer adds workspace-controlled mise or XDG executable directories to its command search path (move tools outside the workspace if you relied on that), and it blocks PERLLIB and PERL5SHELL in workspace environment files. Sandboxed web search now enforces the provider trust policy, so an excluded third-party provider needs replacing with a bundled or verified official one.
Update notes. Updates now check PostgreSQL backup ownership before replacing the installation; if your installed updater lacks the check and an unmanaged Gateway still owns the database, stop that Gateway through its service owner first. Several fixes only apply once the installed updater contains them, and the notes say to update once with your package manager if a 2026.10.7 updater refuses before staging. Browser snapshots no longer expose filled passwords, one-time codes or card fields to the model, and direct skill updates can no longer replace a referenced API key with plaintext. 2026.11.10-beta.1 is a prerelease we do not count.
Release notes →
Affects: /openclaw/, /openclaw/setup/, /openclaw/security/, /openclaw/troubleshooting/
Computer use (beta) is rolling out to Team and Enterprise plans (2026-11-20): Claude can see a user's screen and work in the desktop apps they allow, in the new Claude experience in Claude Desktop and on the web and mobile where "Computer access" is on. Claude Desktop now has release channels (2026-11-17): Stable is the default and updates less often after more testing; Latest gets a new version about once a day. A refreshed Claude Mobile look is rolling out from iOS (iOS 26 or later). On the API, Linked billing is in beta (2026-11-17, user-profiles-billing-2026-10-01 header): customers can pay for the Claude usage in your app through their own Claude organization, and you set billing_policy per request to choose who pays.
App notes →
Affects: /claude-cowork/, /claude-cowork/setup/, /claude-cowork/vs-api/
v0.23.3 (2026-11-19) rolls up about 370 merged pull requests since v0.23.2 into a stable release for the CLI and installers, the Microsoft Store app, Docker and Hermes Cloud, and again defers its curated notes to v0.24.0. The one paragraph it gives names, without detail: a single prompt-cache lifecycle controller (warming through long tool runs, compacting idle chats before the one-hour cache expires, prompt_caching.cache_ttl defaulting to auto), compaction that archives the tool results it drops, the proactive tool-result prune on by default, a one-question first-run CLI setup with the free tier included, and GPT-6.2 Astra and Sol on OpenRouter and Nous Portal. We have not checked those names against Hermes's docs. Release candidates and canary tags for 0.23.3 are not counted.
Releases →
Affects: /hermes/, /hermes/setup/, /hermes/memory/
2026-11-20
Kilo Code
7.8.27 → 7.8.32 — Codestral becomes the default autocomplete
Kilo Code 7.8.32 (2026-11-20) is the new stable release, and it carries the changes from the 7.8.30 and 7.8.31 pre-releases. Mercury Edit 2 autocomplete and Next Edit are retired: Codestral is now the default autocomplete model, and saved Mercury or Inception selections switch to Codestral. The SDK drops client.kilo.edit, and @kilocode/kilo-gateway no longer exports ./edit or ./edit-prompt. You can now change the Kilo telemetry preference from VS Code's About tab; telemetry stays off if Kilo config, VS Code or the environment disables it. 7.8.32 itself stops held keys and repeated close commands from confirming worktree deletion in the Agent Manager.
Releases →
Affects: /kilocode/, /kilocode/setup/, /kilocode/models/
2026-11-18
NemoClaw
0.0.144 → 0.0.145 — onboarding recovery, Hermes routing verification
v0.0.145 (2026-11-18) keeps the selected provider and model when NVIDIA Endpoints onboarding resumes (explicit selections override saved ones), and lets verified resume recovery reuse recorded native-provider credentials when the host lacks the inference key. Hermes routing verification now requires fresh host preflight evidence for provider bindings, effective policy, protected configuration and gateway process identity; missing or changed evidence blocks verification without changing provider state. Generated Switchyard clients use a 120-second deadline per provider call with no automatic retries and no caller credential forwarding. A v0.1.0-rc.1 tag also exists; it is a release candidate we do not count.
Release notes →
Affects: /nemoclaw/, /nemoclaw/setup/, /nemoclaw/switching-providers/
Not counted as news
Prereleases: OpenClaw 2026.11.10-beta.1, Hermes 0.23.3 release candidates and canary tags, NemoClaw v0.1.0-rc.1. OpenRouter catalogue rows for GPT-6.2 Astra Pro and Sol Pro are not on OpenAI's own pricing page, and a Pixtral Large 2502 row is an older model newly listed there. xAI's models page carries a daily date, and the Grok Bot docs overview changed without a dated announcement. OpenAI also shipped GPT-Live-1.1 for its realtime voice API on 2026-11-19, which no agent platform we cover depends on. An OpenClaw blog post titled "OpenClaw Isn't Free." is commentary, not a release. IronClaw has had no release since 1.4.1.
See all releases
Browse the full changelog index for the complete history across all platforms, or the daily one-liner for the most recent state of each agent.