OpenClaw 2026.9.2: New Models and Flexible Workspaces

The 2026.9.2 release is a bridge between the large interface changes in 2026.8.1 and the recovery work that follows. It adds support for GPT-6 Astra and Meta Muse Spark 1.3, gives task runs more flexible workspaces, and focuses on reliability when a Gateway or provider is under stress.

Model names are not configuration

A model appearing in a catalog does not tell you which account, endpoint, price, context limit, or tool set your session will use. The model provider reference keeps provider configuration separate from chat-channel configuration.

Before switching a real workflow, make a small test session. Record the provider, model reference, reasoning setting, and expected output. Then compare the response with the old model before changing a default.

Workspaces follow the task

Flexible task workspaces are useful when a session moves from a local checkout to another placement. They also create a new place to check permissions. Confirm the working directory, branch, and files available to the agent before authorizing a command that writes or publishes.

openclaw gateway status
openclaw doctor

Reliability is part of the feature

The release notes pair model coverage with reconnection and recovery work. That is the right order for an operator: a new model is only useful if a failed request returns a readable error and the session can resume without duplicating an action.

Read the full release notes, then run a bounded test in the same environment as your Gateway. Keep secrets and private transcripts out of test prompts.