Fast tier#

Some vendors offer a paid high-speed tier for some models. OpenProgram surfaces it as a "fast" toggle in the chat UI: when enabled, requests carry the vendor's fast-tier parameter.

Which models have it#

Only two families have a fast tier:

Other models (Gemini, DeepSeek, Qwen, etc.) have no fast concept, and the toggle does not appear for them.

Detection is per model, via the Model.fast field in the runtime registry. A model entry that carries an explicit fast value in config wins — for openai-codex that value is persisted from the official models endpoint's service_tiers data (updated by Fetch after a subscription login, no hand-written list). Entries without one are backfilled from a built-in declaration covering exactly the two families above, matched by model id regardless of provider — the same model resold through a gateway keeps its fast tier.

How to enable it#

• The "fast" menu item / chip in the chat input box: applies per request to the current model. Switching to an unsupported model hides the toggle and the parameter is never sent.
• service_tier in the agent configuration: stores a default tier for an agent, overridable on any turn.

Which providers it applies to#

On the request-building side, only these paths pass the fast-tier parameter through: openai_responses, openai_completions, and openai_codex (service_tier in the request body), plus anthropic (switching to speed: "fast" + the beta header only when the model declares fast support). Other providers do not pass it through, and the parameter never leaves the machine.

Billing note: the fast tier is pay-as-you-go. On a Claude subscription account with no usage credits topped up, Anthropic returns 429 "Usage credits are required for fast mode" — an account issue, not a lack of model support; the UI shows the error as is.

For the full record of implementation details and detection rules, see the design notes.