GPT-6.1 Sol in Codex
OpenClaw v2026.9.9 runs GPT-6.1 Sol in native Codex sessions
OpenClaw's newest release updates its managed Codex app-server and the model catalog bundled with it. GPT-6.1 Sol was already listed, and native Codex sessions can now discover and run it.
Claude Haiku 5.5
The haiku alias now means Claude Haiku 5.5
1M-token context. Haiku 4.5 stays available by its full name.
Claude Haiku 5.5 arrives too, over the API or a Claude Code login. The haiku alias and Anthropic's default utility model now use it. Name Haiku 4.5 explicitly to keep it.
02: Gateway and updates
The Gateway and updates.
Model switches with large plugins
Switching models with large plugins takes about a second
With large plugins installed, changing models used to freeze the Gateway for over a minute, then force a recovery restart. Model runtimes now borrow the Gateway's loaded plugins, so a switch takes about a second.
Background work
Background work stops stalling the Gateway
Skill Workshop reviews and rooted cron jobs used to stall the Gateway as they started. Both now reuse the runtime the Gateway already prepared, so it stays responsive.
Node 24 and 26 exits
Processes exit cleanly on Node 24 and 26
On Node 24 and 26, the OpenClaw CLI, the hook relay and the Gateway could hang after printing their output. They now exit when they finish.
Database ownership check
Updates check database ownership without copying the database
Previous updater stays compatible.
Before updating, OpenClaw checks which process owns its SQLite database. It now inspects the write-ahead log instead of copying the busy database, and the previous release's updater still works.
Update fixes
Updates finish on large installs and FUSE drives
Large npm installs can finish updating again. On a FUSE-backed state directory, backups pass when only file timestamps drift. And Doctor's archive repairs restore the managed Gateway instead of leaving it stopped.
03: Windows and channels
Windows and channels.
Native Windows
Native Windows gets sessions, workspaces and the sandbox back
On Windows, sessions now bind to the physical SQLite store, so stores behind a directory junction work. Empty workspaces initialize natively, and the MXC sandbox runs commands again.
iMessage, Discord, Telegram
Replies reach iMessage and Discord threads again
Ordinary final replies in iMessage go out again. In Discord, subagent results posted into a thread now count as delivered. And Telegram session dashboards launch separately from the owner-only Control UI.
Google Chat, Slack, Feishu
Replies stay in their thread on Google Chat and Slack
Google Chat replies to a threaded message now stay in that thread. So do Slack message-tool replies on CLI runtimes. And Feishu posts now hand their attached files to the agent along with the caption.
04: Security
Security.
Gateway token
Doctor stops writing your Gateway token into config as plain text
Unattended runs skip the step and warn.
Doctor's service repair used to copy a recovered Gateway token into the OpenClaw config file as plain text. It now keeps the token in the secret store and writes only a reference. Unattended runs skip that step and warn.
A2A peers and node policies
Unresolved A2A peer tokens now fail closed
before_agent_run policiesWhen an A2A peer's token variable was unset, the literal placeholder became its credential, and callers got in without a secret. Those peers are now refused. Node turns also enforce your before agent run policies.
05: Before you upgrade
Before you upgrade.
Android and Telegram
This release ships without an Android APK
The Android build was skipped, so the Android app stays on its earlier version. The release owner also waived the Telegram integration checks, so Telegram's package and end to end tests did not run.
Release notes
Every fix and its pull request is in the OpenClaw release notes
The OpenClaw release notes list every fix in this release with its pull request. They're at docs.openclaw.ai, under releases, and on the GitHub release page.






















