OpenClaw Latest Version: 2026.11.2, What Changed and Is It Safe to Update

The latest OpenClaw release is 2026.11.2, published 8 November 2026. It is a patch that fixes update repair blocking itself, scheduled reminders disappearing and duplicate Canvas widgets. If you are on 2026.9.x, read the 2026.10.10 section before you update: it changed fallback for scheduled automations, and large sessions can break it.

Three stable releases shipped in six days: 2026.10.10 on 3 November, 2026.11.1 on 5 November and 2026.11.2 on 8 November. Version numbers on this page were checked against GitHub on 10 November 2026.

What is the latest version of OpenClaw?

2026.11.2, published on 8 November 2026 and marked Latest on the project's release page.

The project describes 2026.9.36 as "a gateway-only extended-stable release, which is our current equivalent to LTS": OpenClaw from the end of September 2026, plus critical security updates, reliability and performance fixes.

What's new in 2026.11.2

The project's own summary: "This is a patch release with no new features. It fixes problems that could stop updates, leave scheduled reminders missing, or cause chat widgets to appear twice, alongside smaller reliability fixes." It draws on 113 pull requests.

The three fixes most people will notice:

• Update repair no longer blocks itself when the current and older install paths reach the same database through different aliases (#193431).
• Scheduled reminders stop disappearing when their text mentions heartbeat polls or wakes (#129650).
• No more duplicate Canvas widgets and previews as a chat reply finishes.

One more for anyone with older scheduled jobs: Doctor now repairs jobs that need a stored inference account by recording an eligible one. After a manual source upgrade, run openclaw doctor --fix before startup.

What changed in 2026.11.1

2026.11.1 landed two days after 10.10. From its highlights:

• Channel replies are more reliable. Telegram, Discord and webhook conversations keep reaching the agent after request handlers or earlier turns finish.
• Update recovery gets unstuck after crashed or canceled updaters, and the Windows Gateway comes back after a failed verification.
• Safer Mac settings. Failed config saves no longer overwrite secrets, and turning off Tailscale exposure keeps Gateway authentication intact.
• Conversations survive more interruptions. Oversized turns keep useful history, and stopping the first Claude CLI turn no longer starts the next message from scratch.
• Doctor migrations report failure instead of success when required work was skipped. If you script around Doctor: per the 2026.11.1 notes, Doctor now exits nonzero when inspection is blocked, the final config is invalid or required migrations are deferred, and scripts using openclaw doctor --json should handle that exit status. For ordinary findings, --json still exits successfully after its report; use openclaw doctor --lint --json to fail on findings.

What changed in 2026.10.10, and what it broke

2026.10.10 is the big one: 3,028 pull requests from 454 contributors. In its own words, it "brings Chat Details beside the conversation, an Advisor that can suggest course corrections during a task, and split speech mode for browser Talk." It also adds event-triggered automations, control from ChatGPT and Claude, and a long list of fixes.

Three things in it deserve your attention before you update from 9.x.

Scheduled automations stay on one inference account. A new isolated job keeps the model account it was created with (#172590). In practice, cron jobs no longer fall back to a model on a different provider. A job that tries fails with:

Automation candidate is incompatible with its recorded inference account and route.

This was deliberate, not a bug. Fallback within the same provider, and interactive chat, still work. A request to allow explicit cross-provider fallback for cron is open as issue #192009. Run openclaw doctor --fix on 11.2 so older jobs get an account recorded.

Large sessions can crash the database worker. The SQLite worker that loads session history has a fixed 512 MB memory cap. A very large session (the report had about 127,000 events and a 1.06 GB database) kills it repeatedly, and every turn in that session fails with Agent database execution admission is closed. As of 10 November, issue #191847 is open and 11.1 and 11.2 do not fix it. The workaround is a new session for the same agent; our admission is closed decoder page has the steps.

Doctor rewrites claude-cli model refs. Doctor rewrites claude-cli/<model> to anthropic/<model> and pins agentRuntime: claude-cli. If an agent ran the same model through both the API and Claude CLI as a fallback, that fallback gets collapsed. This is not new in 10.10 (it has been in the migration since May), and the fix merged to main on 5 November is not in 11.1 or 11.2. Configs already migrated are not repaired automatically: restore from openclaw.json.pre-update or a backup if you lost a fallback.

The error strings people report after updating to 10.10, each with its own page:

• Agent database execution admission is closed
• This Codex session changed before your message could run. Please send it again.
• Codex cannot verify the owner of this model request
• Refusing to run automatic gateway startup migrations

Should you update, and from which version?

• On 10.10 or 11.1: update to 11.2. It is a patch release with no new features.
• On 2026.9.x or 10.x before 10.10: you can go straight to 11.2, but take a backup first and check two things. Do any agents have one very long-running session? Do any cron jobs rely on fallback to a different provider? If yes to the first, plan to start fresh sessions; if yes to the second, read the fallback note above.
• Running a server or a team Gateway: extended stable (2026.9.36) gives you security and reliability fixes without the 10.10 changes.

Older releases cannot open newer database schemas, so the backup is what makes a rollback possible.

How to check which OpenClaw version you're running

openclaw --version

That reports only the CLI on your shell's PATH. If the Gateway runs as a service, it may be a different install, so check the versions in use there too:

openclaw gateway status
openclaw update status --json

Compare the result with the table above. Don't trust a version number you remember installing, because automatic updates may have moved you since.

How to pin an OpenClaw version

To install one exact version, stop the Gateway and install it by number. From the docs:

openclaw gateway stop
npm install -g openclaw@2026.10.9 --allow-scripts=openclaw
openclaw --version
openclaw doctor --fix

Leave out --allow-scripts on npm 11.15 and earlier. For a one-off target without changing your channel, openclaw update --tag <version> updates to that version only; your saved update.channel still decides later updates. To change the policy itself, use --channel stable|extended-stable|beta.

On a Mac, skip 2026.9.6: its app crashed at launch on Apple silicon. The changelog below has the fix.

How to update OpenClaw safely

openclaw update

Before you run it, do three things. Take a backup, because an update that goes wrong at 2am is much easier to undo with one. Note the version you're on now, so you know what to roll back to. And read the release notes for the versions you're skipping, since defaults sometimes change between releases.

That last point matters more than it sounds. 2026.9.2 turned on Swarm sub-agents and opened cross-agent session visibility by default, which changes what your agents can see and do without you editing a single config line. The full procedure, including rollback, is in how to update OpenClaw.

After updating:

openclaw doctor --fix
openclaw gateway restart

Doctor migrates configuration that moved between versions and repairs common breakage. Run it before you start debugging anything by hand.

An update that prints "success" has only proven that the package installed. Whether your config still validates and the Gateway came back are separate questions, and openclaw doctor --fix answers both.

Which release channel should you be on?

Stable is right for most people. You get fixes quickly, and the cadence is fast enough that bugs rarely live long.

Extended stable is for anyone running OpenClaw as infrastructure: a server, a team Gateway, anything where a surprise behaviour change costs more than a missing feature. The docs describe it as a trailing, supported-month release. It's package-only, it never applies automatically, and it gets security and reliability fixes rather than the new features landing on stable. Switch with:

openclaw update --channel extended-stable

Beta exists if you want to help catch problems before they reach stable. Don't run it on the Gateway your business depends on.

Here's the honest framing. The 2026.9 line alone ran to nine stable releases, 9.1 on 3 September to 9.9 on 8 October, and the 10.x line ran to ten more, 10.1 on 10 October to 10.10 on 3 November, before 11.1 and 11.2 followed within a week. That speed is a feature and a cost at the same time.

If reading a changelog every few days isn't how you want to spend your week, that's exactly the gap we built BetterClaw for. Updates are staged and applied for you, and there's a free plan if you just want to see an agent run without babysitting a Gateway.

OpenClaw 2.0, and where it sits

People search for "OpenClaw 2.0" expecting it to be the latest version. It isn't, quite.

OpenClaw 2.0 is the name for v2026.8.1, published on 31 August 2026. OpenClaw's own release notes list it as "v2026.8.1 (AKA OpenClaw 2.0)". It was the big one: a rebuilt web experience, simpler onboarding that detects existing AI keys and local models, conversation search, and sessions that can run on paired devices or cloud workers. Its release scale was 16,977 pull requests and 987 contributors.

Everything since, the 8.x maintenance line, the 9.x and 10.x releases and now 11.x, builds on that. So 2.0 is the current generation, and 2026.11.2 is the current version of it.

If the update breaks something

The most common outcomes, in rough order of frequency:

The Gateway won't start. Run openclaw doctor --fix, then openclaw gateway restart. If it still won't come up, the cause is usually an incomplete install, a config key that moved, or a supervisor entry pointing at the old install. Our Gateway won't start after updating guide maps each log line to a fix.

Settings or skills look reset. Newer releases preserve these through automatic updates, so check that the update actually completed with openclaw update status --json before assuming it wiped your setup.

An agent behaves differently. Check the release notes for changed defaults before you go looking for a bug in your own config. Swarm and session visibility are the two recent examples.

Our OpenClaw security checklist is worth a pass after any update that changes defaults.

Changelog

The sections above cover 2026.10.10 onward. These are the condensed notes for the 2026.9 stable series, newest first, checked against the project's GitHub release notes. If you're updating from any of them, the rule is the same: back up first.

2026.9.7 to 2026.9.9: the cleanup

2026.9.7 (2,818 pull requests) is the one that matters. It carries the source fix for the 9.6 Mac crash ("address async-sleep paths implicated in Mac launch crashes", #156881) and fixes for upgrading from 2026.9.5. It also added verified backups that cover more agent databases, automatic database rollback for eligible updates that fail before the new version starts, OpenAI's Agents API plugin, and Sign in with ChatGPT as a beta.

2026.9.8 is small: repairing an update keeps the plugins you allowed and enabled, and starting directly or in a container can no longer run two copies on the same saved data.

2026.9.9 (112 pull requests) adds GPT-6.1 Sol to the Codex model list and Claude Haiku 5.5 through Anthropic and Claude CLI, and improves recovery after failed updates.

2026.9.6: the Mac crash, and the update that always rolled back

Strip away the incident and 9.6 was a big release: 2,614 pull requests, Claude Opus 5.5, GPT-6 Sol and Luna, Grok 4.7, recovery for unfinished work after restarts, a GitHub reader beside chat and full 30-day usage reporting. One note can bite: if you run a restrictive plugin allowlist, add and enable github to keep the reader and previews.

Then the Mac app aborted at launch on Apple silicon with EXC_CRASH (SIGABRT) (issue #156861). On 24 September the release assets were replaced at 09:52 UTC with a rebuilt, notarized 9.6 DMG, and 9.7 shipped the source fix. The npm package, CLI and Gateway were never affected.

If your Mac app is stuck on the crashing build, it can't update itself:

1. Quit any half-open OpenClaw process.
2. Download the current macOS DMG for your chip from the GitHub releases page. 9.6 was the first release with architecture-specific Mac builds, so pick arm64 on Apple silicon.
3. Drag it over the old app and launch.
4. Run openclaw --version on the Gateway host to confirm the Gateway matches. The Mac app updates a local Gateway it manages, but never a remote or externally managed one.

The quieter 9.6 bug hit Linux servers. Managed npm updates from 9.5 under systemd failed in update-history reconciliation and rolled back (issue #157011), with:

Update history reconciliation could not complete: RangeError: Maximum call stack size exceeded

The rollback was the safety net working: Gateways stayed healthy on 9.5. doctor --fix didn't clear it, and pruning update history by hand was discouraged. 9.7's fixes for upgrading from 9.5 are the answer.

2026.9.3 to 2026.9.5: updates get safer

2026.9.3 rehearses core and plugin changes in isolated candidate state before activating them. It also added shareable conversations: a revocable, read-only public link to a session's text, current and future. Treat it like any link anyone can open.

2026.9.4 can put the previous app version back when a compatible update fails, but rolling back the app does not restore your data, so you still need your own backup. It also put bundled and ClawHub plugins into one Plugins workspace and added OPENCLAW_CONFIG_READONLY=1 for deployment-managed config.

2026.9.5 brought Atomic Updates that check the next version before switching over, and plugins you can install without restarting the Gateway. It's also the release the maintainers pointed Mac users back to during the 9.6 crash.

2026.9.2: two defaults flipped

The line that mattered most wasn't in the Highlights. Under Changes: "session tools now default to all-session visibility and ordinary agent-to-agent access is enabled." If you run more than one agent on a Gateway, each can read the others' sessions unless you set tools.sessions.visibility to agent or self. Do that on any agent that handles credentials, money or private messages. Our OpenClaw multi-agent setup guide covers why isolation matters.

Swarm is enabled by default, too: concurrent sub-agents with structured results and live progress, with explicit opt-outs, tool restrictions and the separate Code Mode opt-in preserved. A single bad instruction can now be obeyed by several workers at once.

# for each agent that handles anything sensitive
tools.sessions.visibility = "self"
# and decide, deliberately, whether swarm stays on

9.2 also added openai/gpt-6-astra (OpenAI API-key profile or an eligible ChatGPT/Codex account), with /think ultra orchestration. Before making it your default, read our GPT-6 Astra vs Fable 5.1 vs Sonnet 5 test: it was the most expensive per task by a wide margin. The reliability fixes were overdue: automatic updates keep active settings, enabled skills and default-agent ownership, active, queued and delegated replies survive a Gateway restart, and backups reject corrupt archive headers instead of accepting a partial backup.

Do you need to update every time?

No, and opting out of chasing every release is reasonable. Update promptly for security fixes, update deliberately for features, and consider extended stable if "did anything change in this release?" is a question you don't want to ask every few days.

The version number is the easy part. Knowing what changed underneath it, every week, on a Gateway other people rely on, is the real job. Some teams love that job. Plenty would rather spend the time on what their agents actually do, which is why managed platforms exist as a category. If you're weighing that trade, our guide to AI agent builder platforms lays out the options.

If that's you, give BetterClaw a try. The free plan gives you one agent and 100 credits a month with no card. Starter is $19 a month, and Pro is $49 a month for five agents, all with your own model keys and no inference markup. There's no changelog to read and no 2am Gateway that won't start. Start free or see full pricing.

Frequently Asked Questions

What is the latest version of OpenClaw?

OpenClaw 2026.11.2, published on 8 November 2026, is the current stable release. It is a patch release on top of 2026.11.1 and the large 2026.10.10 release. The extended stable (LTS) line sits at 2026.9.36, published on 29 October. Check your own version with openclaw --version, and openclaw gateway status for the Gateway service.

Is OpenClaw 2026.11.2 safe to update to?

Yes for most setups, especially from 2026.10.10 or 2026.11.1: it is a patch release with no new features. Coming from 9.x, take a backup first and check for very large sessions and cron jobs that fall back to another provider, because 2026.10.10 changed both. The large-session database crash (issue #191847) is still open in 11.2.

What changed in OpenClaw 2026.10.10?

It added the Chat Details panel, an Advisor that suggests course corrections during a task, split speech mode for browser Talk and event-triggered automations, across 3,028 pull requests. It also binds each scheduled job to one inference account, which stops cross-provider fallback for cron jobs.

How do I pin an OpenClaw version?

Stop the Gateway, then install the exact version: npm install -g openclaw@2026.10.9 --allow-scripts=openclaw (drop --allow-scripts on npm 11.15 and earlier), then run openclaw doctor --fix. For a one-off update to a specific version, use openclaw update --tag <version>. Rolling back onto an older version needs a pre-update backup, because older releases cannot open newer database schemas.

Why did my fallback model stop working after the update?

Since 2026.10.10, scheduled automations keep the inference account recorded for the job and refuse fallback to a different provider, with "Automation candidate is incompatible with its recorded inference account and route." This is intentional. Fallback within the same provider and in interactive chat still works, and a feature request to allow explicit cross-provider fallback for cron is open as issue #192009. If an agent lost a claude-cli fallback, Doctor's model-ref rewrite collapsed it: restore it from openclaw.json.pre-update.

Is OpenClaw 2.0 the same as the latest version?

Not exactly. OpenClaw 2.0 is the name for v2026.8.1, published on 31 August 2026, which brought the rebuilt web experience and simpler onboarding. Every release since builds on it, so you're on the 2.0 generation, and 2026.11.2 is its current version.

How do I update OpenClaw to the latest version?

Run openclaw update, then openclaw doctor --fix and openclaw gateway restart. Take a backup first and note your current version, so a rollback is straightforward. If your Mac app is stuck on a crashing 2026.9.6, download the current DMG from GitHub by hand, because the app can't update itself.

Is it worth updating OpenClaw every time a release ships?

Not always. Stable shipped more than twenty releases between 3 September and 8 November 2026, so following every one is a real time cost. Update promptly for security fixes and deliberately for features. If you want fewer, quieter updates, extended stable is a slower line that never applies automatically.

Is the latest OpenClaw version safe to run on a production Gateway?

Stable is fine for most setups, but it moves fast and occasionally ships a bad build, as 2026.9.6 did on the Mac. For a Gateway other people depend on, extended stable (2026.9.36 today) is the safer choice, because it only gets security, reliability and performance fixes. Whichever you run, back up before every update.