Get a suggestion

POSThttps://api.inferencebureau.com/v1/suggest

Pre-flight confirmation for mixed-model sessions. Send the turn you are about to send; get back the least expensive pool model that meets the project's quality floor for it, with a token. Send the turn with ib.suggestion set to that token to run it on the suggested model, or without it to stay on the requested model.

Body

application/json

modelstringrequired
The model you would request for this turn. Suggestions are issued against it, and the turn must request the same model to use the token.
messagesarrayrequired
The turn exactly as you will send it. A Messages body (with system or Anthropic content blocks, or an anthropic-version header) is read as a Messages turn, and only suggests models that endpoint serves.
input, instructionsstring | arrayoptional
Instead of messages, for a Responses turn: the body you will send to /v1/responses. It is read as a Responses turn, and only suggests models that endpoint serves. max_output_tokens is used for the estimate.
toolsarrayoptional
Sent so only models that support tools are suggested.
response_formatobjectoptional
Sent so only models that support structured output are suggested.
max_tokensintegeroptional
Used for the cost estimate. max_completion_tokens is accepted too.
userstringoptional
Your end user, as on the turn (metadata.user_id in a Messages body). Recorded on the classifier call for attribution and per-end-user Directives, and not forwarded.
ibobjectoptional
session groups the call with its conversation. task skips the classifier and is not charged. tags tags the classifier call, as on the turn. See the ib object.

Response

suggestionstring | nullrequired
The token to send as ib.suggestion. null when no cheaper model fits; then send the turn as usual.
modelstringrequired
The model the turn would run on: the suggested one, or the requested one when there is no suggestion.
taskClassstring | nullrequired
One of extraction, summarization, rewriting, translation, general_qa, coding, complex_reasoning, unknown.
confidencenumber | nullrequired
Classifier confidence, 0 to 1.
reasonstring | nullrequired
Why this was or wasn't suggested, e.g. cheapest_passing, no_passing_model, low_confidence. See reasons.
estimateobjectrequired
Estimated USD for the turn on the requested model (baselineUsd) and on the suggested one (suggestedUsd), and the classifier charge (classifierUsd). Estimates use a typical reply length.
expiresIninteger | nullrequired
Seconds until the token expires. Pre-flight tokens last 2 minutes.
requestIdstringoptional
The receipt for the classifier call, in Logs. Absent when no call was made.

Headers

Authorizationstringrequired
Bearer ib_live_… for a key set to Suggest under mixed-model sessions.
ib-sessionstringoptional
Same as ib.session.
ib-userstringoptional
Your end user, up to 256 characters; wins over the body's. See End users and tags.
ib-tagsstringoptional
Comma-separated tags, such as inference-bureau-search, at most ten with ib.tags. Never sent upstream.

When it's available

The key must be set to Suggest, and the project must use Pre-flight confirmation. Otherwise:

403 policy_denied when the key is Off or in Shadow.
400 invalid_request when the project confirms inline. Read the suggestion from the chat response instead.

Billing

Each call that reaches the classifier is its own request, billed at Jev's input-token price plus the platform fee, with a receipt. No charge when a tool cycle is open (the response says tool_cycle_open and carries ib-pinned: tool_cycle), when you send ib.task, or when nothing can be sent to the classifier. The classifier call counts against your Directives like any request: per-end-user limits and budgets for the end user it names, and tag budgets for its tags. endUser.required refuses a call that names no end user.

Calls have 3 seconds, including one retry if the classifier is at capacity. A timeout returns no suggestion rather than an error, so the turn can go ahead on the requested model.