Skip to main content
Chat and the Models page now open from the catalog OpenClaw already has, so choosing a model no longer waits for a full provider scan. Live discovery runs only when you open a model screen or explicitly refresh it, keeps the last useful list when a lookup fails, and /model can change only the current conversation or deliberately update one agent or the shared default. Once a model is selected, OpenClaw keeps the request inside the intended provider and authorized account order, preserves the real order and outcome of streamed replies and tool calls, carries reasoning and context settings with the model and runtime, and reports plan windows, token use, context pressure, and estimated cost more clearly.
Chat and the Models page now start from the catalog OpenClaw already has, with supported providers looking for newer chat and text models when you open a picker or request a refresh. If that lookup fails, the built-in entries and last working list remain available.A /model change can stay with the current conversation or deliberately apply to one agent or the shared default, with persistent changes requiring the right authority. Aliases and fallback keep the provider and account attached to the selected model.The list shows models OpenClaw can identify, while the provider, account, region, endpoint, plan, limits, and pricing determine which ones you can use.
OpenClaw now keeps model requests inside the provider accounts and credential order you configured. If one account hits an authentication or quota cooldown, the next authorized account for that provider can take over without changing the selected provider or model, and the saved preference resumes when it recovers. Environment keys remain available when no explicit account list is configured.OAuth registration stays with the setup conversation where it began, failed credential writes surface as failures, and a successful login applies only to the model route that authenticated. Tenant and custom-endpoint credentials stay on the account and origin they were configured for.New GitHub Copilot device logins place the token in OpenClaw’s protected local secret store by default and keep a reference in the auth profile. The store depends on state-directory permissions rather than encryption at rest, existing inline profiles are not migrated, and operators can still choose the prior plaintext mode explicitly.
OpenClaw now distinguishes a complete streamed reply from one that stopped partway through. On supported Responses paths, reasoning, text, tool output, and available usage remain in order; if a stream fails, valid work that already arrived is preserved, and OpenClaw retries or moves to an authorized fallback only when replay is safe. With no safe recovery path, the partial reply remains attached to the error.Tool calls wait for a complete name and arguments before they can run, including large streamed arguments and calls whose provider item IDs change along the way. Unmanaged native OpenAI Responses conversations on the official endpoint with storage enabled can continue by sending only new input after the first turn, then recover with full history if that upstream continuation state expires.
Reasoning settings now stay attached to the selected model and runtime when a conversation is restored or its authentication route is rebuilt. Native Codex supports Ultra for Sol and Terra and Max for Luna; embedded OpenClaw maps Ultra to the provider’s highest supported effort and adds guidance for delegated work, which is a different behavior from native Codex Ultra.Supported GPT and Claude routes expose larger context options. Normal GPT-5.5 and GPT-5.6 runs use a 272,000-token budget with an opt-in 922,000-token input window, and the Control UI can choose 200K or 1M for supported Claude 5 CLI conversations. These options remain limited to the routes, interfaces, and accounts that support them.Compaction now uses the latest trustworthy context count rather than accumulated cache billing or duplicated history, preserves tool output across supported Responses checkpoints, and falls back to full history when a stored checkpoint is rejected. Context limits are now configured per model, and Doctor can migrate supported provider-level settings tied to explicit model entries.
OpenClaw now separates subscription-plan information from estimated API cost. Chat can show plan windows, reset times, credits, and the account email attached to a snapshot, while completed iOS replies can show supported input, output, cache, cost, and context-pressure details. These figures are snapshots or estimates, not provider invoices, and in mixed API-key and subscription setups the account label identifies the plan snapshot rather than every run.The Control UI adds a Profile page for lifetime activity recorded by OpenClaw, with Usage and Profile views grouped in the selected time zone, and plugin-initiated model calls now contribute to aggregate totals. OpenClaw cannot reconstruct activity it never recorded.Supported failed or incomplete turns can retain the provider’s token and cost data without being marked successful. Permanent authentication, model, media, and long-window quota failures stop retrying, while transient rate limits and retryable server errors keep their existing retry or authorized same-provider fallback behavior; Codex subscription runs do not silently switch to pay-as-you-go API keys.