Fast

Some vendors serve a model faster for a higher price. /fast asks the model in force for that speed, where its vendor serves one, and says what it costs before you choose it.

Choosing it

/fast opens a panel over the model in force: its provider, its name and the credential it is served by, what the vendor says of fast, and two rows, Standard and Fast, with the price and the vendor's speed beneath Fast. The mark opens on the speed in force, so Enter changes nothing by accident; Escape leaves it as it was.

Speed · openai · gpt-6-astra · OpenAI key
 
OpenAI may serve a fast request at standard speed when fast capacity is short;
it is then billed at the standard price.
 
  Standard
  The standard price and speed
 
› Fast
  2x the price · up to 2.5x faster
 
enter to choose · esc to cancel

/fast on and /fast off choose without the panel. A window with no room for the panel, and a run with no keyboard, print the speed in force and those two lines, the price beside on. While a turn runs, /fast opens as /model does, and the speed taken is asked for once the turn ends.

A model with no fast form, and a model that is itself a fast model, answer /fast with one line and no panel:

⎿ anthropic · claude-sonnet-5 has no fast form
⎿ moonshot · kimi-for-coding-highspeed is a fast model of its own; /model lists the others

In /model, a model with a fast form says fast at the end of its row, after no rung where it has both.

What you are told afterwards

The label under the box ends in · fast only after an answer the vendor says it served fast. A vendor that serves a fast request at standard speed says so in its answer, and the label then leaves fast off until one is served fast again; nothing is added to the transcript.

Where it is kept

/fast writes the model in force and "fast": true together under the provider in the configuration file in your home directory, since the speed was chosen at that model's price, and /fast off takes fast out:

{ "providers": { "openai": { "model": "gpt-6-astra", "fast": true } } }

The model written is the one in force, even where --model or a project's file chose it, so it becomes the one a plain start asks for where no project file names another, and /fast off leaves it there.

Only that file is read for it. A project file that sets fast stops crucible before it draws anything, as every key a checkout may not set does: fast costs more on every request, and a repository could otherwise spend your money.

A request carries the fast form only where all of these hold: your file says fast for the provider beside the model in force; the model in force has a fast form in the table below for the credential in force; and no baseUrl is set for the provider. A fast left in the file where one of these does not hold is left there and asks for standard. Making room in the model's window asks at the same speed as a turn.

When it goes back to standard

The speed goes back to standard, and fast is taken out of the file, when:

• another model is chosen for the provider in /model. Choosing the model in force again keeps it;
• a credential is stored, removed or replaced for the provider and the one in force changes by it: a /login on another row, a key stored where the environment served, a /logout that leaves the environment serving, or a second credential taken out at a start. What fast costs was shown for the credential it was chosen under. A key written again on the same row, and a change of the environment between two runs, move nothing;
• the vendor refuses the fast form, as below.

When the vendor refuses it

Where a vendor answers a fast request with a refusal of fast before anything else, crucible sends the same message once more at standard speed, turns fast off, and says why in one line:

⎿ openai refused fast: Invalid service_tier argument: The requested service tier is not allowed for this project. Sent again at standard speed; fast is off.

Stopping the turn while the message goes again leaves the line ending Fast is off. instead.

Any other error is reported as it would be at standard speed, and nothing is sent again. What counts as a refusal of fast:

• OpenAI with a key: a 400 whose error is invalid_request_error for the parameter service_tier, as OpenAI's error reference documents.
• OpenAI with a ChatGPT sign-in: none is documented, so none is taken for one.
• Anthropic: a 400 invalid_request_error whose message names fast mode or speed, and a 429 saying "Usage credits are required for fast mode.". Anthropic says only that such a request "returns an error"; this rule is taken from open harnesses that are not Anthropic's.
• Google: none is documented. Congestion does not refuse; it serves the request at standard speed, which the answer then says.
• Kimi Code: kimi-for-coding-highspeed is a model of its own and asks for nothing, so there is nothing to refuse. A plan without it answers with an error like any other, and nothing is sent again.

The fast forms

Each price, caveat and speed is the vendor's, as the panel shows it: the speed beside the price under Fast, where the vendor states one.

OpenAI states twice the price for GPT-6 Astra, GPT-6.1 Sol and GPT-5.6 Sol; for GPT-6 Sol, GPT-6 Luna, GPT-5.6 Terra and Luna, and for GPT-5.5, the multiple is worked out from the prices on its pricing page. The ChatGPT speed page states no speed for the GPT-6 models or GPT-6.1 Sol. OpenAI serves no fast form of the GPT-6 models or GPT-6.1 Sol with EU data residency.

A provider whose row says None in this release serves a faster form, but no source settled how it refuses one, or how an answer says which speed served it; crucible offers fast only where both are known. A row that says None names a vendor that documents no faster form crucible could ask for.