Get a request

A req_… id from your team, from the X-Request-ID header or error.request_id. Any key of the team can read any of its requests. Unknown ids, and ids from other teams, return 400 invalid_request.

chat.completions, messages, responses, embeddings, evaluate or suggest.

The model you asked for, as sent: a team alias stays org/fast.

Only when the alias chose its model through a routing profile: profile, objective, quality_floor, task_class, the model chosen, the reason (cheapest, highest_quality, or baseline when no pool model qualified), the baseline, and candidates: each pool model with its status (selected, eligible, below_floor, not_evaluated, or why it couldn't take the request, such as policy_denied), score and estimated est_usd.

Only when the request named a team alias: name, the model it resolved to, and the defaults its preset filled (body fields such as temperature).

require_zdr and require_no_training: the data Directives the request ran under.

The model that answered: a fallback model or an accepted mixed-model suggestion when one applied. null if none answered.

The provider that served it.

explicit for the model as requested; mixed_accepted or pinned for mixed-model turns; profile when an alias's routing profile chose the model.

The ib-session id the request was sent with.

The end user it was sent for (ib-user, user, safety_identifier or metadata.user_id). See End users and tags.

Its tags (ib-tags and ib.tags); empty when none were sent.

admitted, dispatching, streaming, succeeded, failed, canceled, or unknown.

See request states.

The error code the request ended with.

Each provider attempt in order, with its model, provider, state, upstream status, error, and timing (below). operation is execute for the request itself or classify for a mixed-model classifier call it carried.

Token counts per generation, including reasoning and cached tokens, and the finish reason. Empty until usage is known.

With status: "final": upstream_cost, platform_fee and customer_charge as USD decimal strings, and the fee rate in markup_bps. Otherwise status is pending until usage settles, or released when nothing was billed.

Present when the request produced or used a mixed-model suggestion: the suggestion id, task class, confidence, models and estimates.

Directives, price plan, and catalog versions that applied.

When the attempt was sent to the provider.

For a stream, when the first event arrived, before any byte reached you. For a non-streamed call, when the whole response arrived.

When the attempt finished. For a stream, when its last event was relayed, so a slow reader makes it later.

Whether the attempt streamed. null until something came back.

Time to first token: first_byte_at − started_at. Streamed attempts only; null for non-streamed calls, whose first byte is the whole answer.

ended_at − started_at. null while the attempt runs.

Completion tokens (reasoning included) divided by generation time: ended_at − first_byte_at for a stream, ended_at − started_at for a non-streamed call, which includes the wait for the first token. null until usage is known, or when no completion tokens were reported.