Claude Code adds provider controls and clearer model fallback
September releases improve provider restrictions model fallback and gateway usage accounting
Claude Code’s September 29 and 30 releases give administrators more control over API providers and make model fallback easier to interpret. Versions 2.1.285 and 2.1.286 also correct usage accounting and session recovery behavior, changes that matter when coding agents run through enterprise gateways.
Provider restrictions become a managed setting
The September 29 release introduces the allowedProviders setting, letting administrators restrict which API services a machine may use. Supported choices include the Anthropic API, custom endpoints, Amazon Bedrock, Vertex AI, Microsoft Foundry and other supported deployment routes. The same release adds a control for disabling WebFetch and fixes usage reports that could assign a turn to the wrong model after a refusal triggered an unexpected fallback.
Scope matters when deploying those controls. Claude Code’s settings documentation distinguishes organization policy from personal and project settings. A configuration on an employee’s device does not automatically become policy on a separate cloud machine. Administrators need to check which settings source reaches each environment rather than assume one local file governs every session.
Fallback gets clearer notices and accounting
In version 2.1.286 on September 30, Claude Code retries once on the previous model in the same tier if the Anthropic API rejects the model selected by a default or alias. Refusal retries also switch to standard speed when the fallback model cannot use fast mode. Updated notices explain when a fallback reduces the context window from one million to 200,000 tokens.
The release also corrects the Claude apps gateway spend meter. It had priced one hour prompt cache writes at the cheaper five minute rate and could miss input tokens after the first model call in streamed turns involving a server tool. Other fixes address missing conversation turns after interrupted parallel tool calls and cloud sessions that could not wake with large histories.
Safety routing depends on the model and topic
Current model configuration documentation separates content based fallback from availability fallback. It says cybersecurity requests flagged on Sonnet 5.5 run again on Sonnet 5. Biology requests flagged on that model end in a refusal because it has no biology fallback.
For Fable 5.1, Fable 5 and Opus 5.5, the documented routes are Opus 4.8 for flagged cybersecurity requests and Opus 5 for flagged biology requests. A fallback target blocked by the model allowlist cannot run. These are documented routing rules, not an October 1 model announcement.
What teams should check
Taken together, the changes make the path from a selected model to a completed task more visible. A useful rollout check would record the selected model, any fallback notice, the final usage record and whether the session kept its expected context. That gives teams a way to distinguish an application problem from a provider restriction or a documented model switch.
Cost dashboards deserve a separate review after the update. Corrected metering can change what a dashboard displays without establishing that the provider changed its prices. Teams should reconcile gateway measurements with provider billing before interpreting a difference as a new rate.
Featured image is an original AI generated editorial illustration.
