Upgrading OpenClaw 2026.9.8: GPT-6.1 Sol support limits and delegated replies

First identify the upgrade in question

OpenClaw v2026.9.8 was officially released at 03:21 UTC on October 3, 2026. When reading older tutorials, distinguish the software version, article date and version on your own machine. This article focuses on checks after upgrading an existing installation, using official documentation; it does not claim to have tested your environment.

On October 4 UTC, the official notes for this release were updated to clarify that GPT-6.1 Sol support was incomplete. Upgrading to 2026.9.8 can no longer be treated as a complete way to enable Sol. The following mainly serves package-manager users checking update recovery or delegated-reply fixes. Keep the existing model, validate basic functionality, then assess model changes separately.

Preview first, then update through the appropriate installation method

These direct-upgrade steps target recent versions. For older installations, first check the legacy migration process in the official Updating documentation rather than directly applying these commands. Before starting, create a backup as described in the final section, then run openclaw update status to inspect installation and update status. For global npm, pnpm or Bun installations, preview with openclaw update --tag 2026.9.8 --dry-run, verify the target, then run openclaw update --tag 2026.9.8. --tag sets the target for this package update. If using extended-stable, do not apply this combination directly; read the official channel guidance first.

Source, container and desktop-app users should use their respective update entry points. In particular, a normal exit code is not proof of an upgrade. Official guidance says an installation whose package-manager ownership cannot be identified may report skipped and ask you to update through the original installation method.

Before executing, confirm that you are updating the installation actually running the service, and choose a time when interruption is acceptable. Keep the upgrade preview, original version and installation method. If progress stalls, inspect the current step rather than starting another update from a second terminal.

Check running behavior, not just installation completion

A successful managed openclaw update includes Gateway restart and verification. Follow any remaining service-repair warnings. For diagnostics, run openclaw --version, openclaw health and openclaw gateway status --deep to inspect the command-line version, health and service status respectively.

Compare outputs with pre-upgrade records. If the CLI shows the new version but the service is not running correctly, address service targeting or startup errors before signing into model accounts again. If a simple conversation that previously worked now fails, investigate the base service first. Use a short task with a model that worked before the upgrade; do not make incompletely supported Sol the acceptance criterion.

Sol support is incomplete: inspect the current model first

The latest stable-release notes describe Sol model selection, API-key and subscription routing work as preparatory, explicitly saying support was not completed in 2026.9.8. The selection commands from the old notes have been removed.

Use openclaw models list --provider openai to inspect the catalog, then openclaw models status for the default model, fallbacks and authentication overview. models set changes the global default. To inspect a particular agent, use openclaw models status --agent AGENT_ID, replacing AGENT_ID with your actual identifier rather than copying the placeholder unchanged.

First establish which agent and account handle the current task, and retain a model already validated as usable. Before trying Sol, look for stable-release notes explicitly confirming completed support. This documentation correction gives no completion version or date; a development-branch commit is not a released capability. In multi-user or multi-agent environments, success in another session does not prove that the current session uses the same settings.

A selected model can still fail: distinguish causes using evidence

For a known provider, models set may save a model selection and warn even if that model is absent from the local catalog. Saving successfully therefore does not establish usability. models status without --probe is not a model invocation test either. --probe makes real requests and has operational requirements such as exclusive use of a state directory; do not add it casually.

For other model issues, investigate by evidence: compare model or permission rejection with the account catalog; check the active login credentials for authentication errors; and determine whether a timeout or connection failure concerns the local Gateway or an unfinished external request. This is a diagnostic order, not a root-cause claim based on one error. If the original model works, use it to finish the work and preserve the error for review against a later stable release.

Delegated results now return once; old workflows need explicit follow-ups

For ordinary sessions delegating through sessions_send, the result returns immediately or once to the original requesting session after completion. The old automatic agent back-and-forth and target announcement turns were removed. Continuing discussion or sending content to a channel requires an explicit follow-up request or messaging-tool call. Special paths such as isolated Cron may return skipped; judge by delivery status rather than requiring an additional duplicate delivery.

For example, ask agent A to have agent B list files in a test directory read-only, then observe whether the result returns to A. Next, have A explicitly ask B to explain one file, testing follow-up with two independent requests. Lack of automatic continued discussion is not an upgrade failure, nor does this change mean every external group chat must receive a reply. If your old workflow depends on automatic handoffs, rewrite that step before validating the full task.

Keep a recovery point when upgrades fail; do not overwrite the old package blindly

Before upgrading, openclaw backup create --verify can create a verified backup. Reverting the package version does not undo configuration or database migrations. Preview a specified rollback version through the official process only if the old version can read the current state. Incompatible state requires a software version matching the backup and offline restoration. Do not remove recovery files listed in the update report until installation and session history are confirmed healthy.

Treat “restore service” and “try a new model” as separate tasks. Restore a known-working environment first, then adjust the model independently. If backup restoration is necessary, preserve current state and identify sessions created after the backup so a single model rejection does not erase new work. If a newer format is rejected, do not manually alter version markers to bypass checks; follow the report and official recovery documentation below.