Fast mode extended to Vertex AI, session-only and limited to some regions

What

Fast mode is a setting that makes Claude reply faster at a higher price. It was limited to the Anthropic API directly (the "first-party" provider). It is now also available on Google Vertex AI, a cloud service that hosts Claude models, for models that support it there.

• Which models qualify: model capability lists now include a new vertex_fast_mode entry next to the existing fast_mode. On Vertex, fast mode is offered only when the model's capability data explicitly lists vertex_fast_mode. On the Anthropic API, fast_mode is used, and when capability data is missing Claude Code still falls back to checking the model name for opus-4-8 or opus-5.
• Availability check: fast mode is now available when the current provider is the Anthropic API or Vertex and CLAUDE_CODE_DISABLE_FAST_MODE is not set. Before, any provider other than the Anthropic API was refused outright. One finding notes that some of the surrounding code, including the part that handles CLAUDE_CODE_SKIP_FAST_MODE_ORG_CHECK, may simply have been moved around rather than changed.
• Region requirement: on Vertex, fast mode is turned off with reason disabled_by_env when the model's region is not global, us or eu. The message reads "Fast mode needs Vertex region global, us or eu" and tells you which region setting (an environment variable) to change before restarting. Before, disabled_by_env just said "Fast mode is not available".
• Session-only on Vertex: turning fast mode on or off with /fast now saves the fastMode setting only on the Anthropic API. On other providers the change lasts for the current session, and the confirmation ends with " (this session only)". The settings description now says "On Vertex a saved value is ignored: use /fast each session." A saved value is also no longer treated as "on by default" outside the Anthropic API, even when fastModePerSessionOptIn is not set.
• Opt-in prompt: the stored fastModeChosenOnFableOrMythos flag is now recorded, and stops the opt-in prompt from appearing, only on the Anthropic API.
• Wording: the not-available message changed from "Fast mode is only available when using the Anthropic API directly" to "Fast mode is not available on this provider". The pricing line says "Draws from usage credits" on the Anthropic API and "Billed" elsewhere. Other providers see a price label reading "2x ... standard price". On Vertex, the tip says fast mode switches to "a model that does" instead of naming Claude Opus.
• When a provider rejects fast mode: on providers other than the Anthropic API, an error with status 403 or 404, or a 400 whose message mentions fast mode, "speed: extra inputs" or the speed parameter, now makes Claude Code turn fast mode off and retry the request. A 400 marks that model as rejecting fast mode, and other statuses start a waiting period. The notice now reads "Fast mode rejected for [model]; using standard speed for it this session". The words "fallback model" were dropped, so the notice is no longer only about the backup model Claude Code switches to.

Why

If you use Claude through Vertex, you may now be able to use fast mode on supported models in the global, us or eu regions. Expect to run /fast again in every session, because a saved fastMode value is ignored on Vertex. If Vertex refuses the faster speed, Claude Code drops back to standard speed and retries instead of failing the request.