Request types

There is one model. Every call is classified into a request type by its shape, and each request type invokes the model best suited to it.

How a call is classified

Classification is deterministic and local — it reads the request shape (input size, tools, schemas, code, images, documents, reasoning language) and combines it with the intent implied by your model id. Capability signals win over the model name: a PDF sent to a “mini” model is still a document request. Rules are evaluated top to bottom; the first match wins.

#	When the request has…	Resolves to
1	Document / PDF / file input	document
2	Image input	vision
3	Input estimated beyond the standard context window	long
4	Tool definitions or a forced tool choice	code
5	Code blocks, diffs, stack traces, or repo paths	code
6	A JSON / response schema, or extraction language	json
7	Multi-step reasoning language	reason
8	Simple summarize / classify / translate / rewrite	flash (or pro at higher effort)
9	General chat with no special signal	the model-intent base type (pro or flash)

The request types

These are the request types your traffic can land on. You see the request type in your own analytics; the model, candidate, and provider that serve it stay private.

Request type	Triggered by
vision	Image content parts in the request Requests with image input.
document	PDF / file input (e.g. Anthropic document blocks) Requests with PDF / file input.
long	Estimated input beyond the standard context window Requests whose input exceeds the standard context window.
code	Tool definitions, diffs, stack traces, or repo paths Code-shaped requests — diffs, stack traces, tool use.
json	response_format / output JSON schema set Requests that set a JSON / response schema.
reason	Multi-step reasoning language in the prompt Multi-step reasoning requests.
flash (default)	Simple request at low effort Simple, low-effort requests — served fast and cheap.
pro (default)	Everything else (default) General requests with no special signal.

Examples

Request	Resolves to
General text on a Claude / GPT id	pro
Code or tool calls on any id	code
Summarization on a small / mini id	flash
Any id with a PDF document block	document
Any id with image input	vision
Any id with a JSON response schema	json

You never select a request type

The request type is inferred from the call — there is no parameter to pick a type. The only lever you control is effort, which biases cost, latency, and quality within the chosen type without changing your code.

See the type your calls land on

Each response returns its type in the X-DI-Request-Type header (see Response headers), and the mix across all your traffic is broken down in Usage & analytics.