Chat Completions
The standard OpenAI Chat Completions API on every catalog model, with what is honored, dropped or refused.
Send a request
Standard OpenAI body; model must be a slug from GET /v1/models. The response adds usage.cost (USD charged), usage.is_byok, and a top-level provider naming who served it.
/v1/chat/completionsGET/v1/models/{model}curl https://api.experientiallabs.ai/v1/chat/completions \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-6-luna",
"messages": [
{"role": "system", "content": "Be brief."},
{"role": "user", "content": "Name three prime numbers."}
],
"max_tokens": 256
}'Fields: honored, dropped, refused
Every field gets one of three treatments. Nothing is rewritten silently.
| Treatment | When | Signal |
|---|---|---|
| Honored | The serving provider supports it | none |
| Dropped with disclosure | The provider cannot honor it but the request is servable (top_p on a non-sampling route) | top-level x-experiential-ignored-parameters array in the body (on a stream, in the finishing chunk) |
| Refused | No route serves it, or it is malformed | 400 unsupported_parameter or invalid_parameter, param names the field; nothing charged |
Refused on every route today: audio, modalities, logit_bias, moderation, prediction. seed and prompt_cache_retention are dropped with disclosure. The deprecated functions / function_call are rewritten to tools / tool_choice with disclosure.
{
"id": "chatcmpl-...",
"provider": "openai",
"x-experiential-ignored-parameters": ["web_search->dropped(search_unavailable)"],
"choices": [...]
}Reasoning effort
reasoning_effort passes through when the route supports it. An unsupported level may move down (never up) to the closest supported one when the route's policy allows; otherwise 400. none means reasoning off and can be refused; minimal still reasons; max is the top tier. Substitutions appear in x-experiential-ignored-parameters.
curl https://api.experientiallabs.ai/v1/chat/completions \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "gpt-6-astra", "reasoning_effort": "high", "max_tokens": 4096,
"messages": [{"role": "user", "content": "Prove there are infinitely many primes."}]}'Web search
Works on every model. Ask with web_search_options, OpenRouter's plugins: [{"id": "web"}], or a :online model suffix. Routes with native search use it; elsewhere the gateway runs one search for the latest user turn and cites sources as url_citation annotations. Each gateway-run search costs a flat fee (default $0.007, up to 10 results, never on BYOK), included in usage.cost; the count is usage.server_tool_use_details.web_search_requests.
curl https://api.experientiallabs.ai/v1/chat/completions \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "gpt-6-luna", "web_search_options": {},
"messages": [{"role": "user", "content": "What changed in the latest Node.js release?"}]}'
# citations: choices[0].message.annotations[] with type "url_citation"Safe retries with Idempotency-Key
Delivery is at least once: retrying an ambiguous failure can bill twice. Send an Idempotency-Key header and an exact retry replays the original bytes instead of running again. Same key with a different body: 409 idempotency_conflict. Replay lives in the serving worker's memory, so it is not guaranteed after a restart. Keyed responses carry no usage.cost; read cost from usage. Not honored on embeddings or Messages.
curl https://api.experientiallabs.ai/v1/chat/completions \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Idempotency-Key: 7c1d6a3e-order-1234" \
-H "Content-Type: application/json" \
-d '{"model": "gpt-6-luna", "messages": [{"role": "user", "content": "Summarize order 1234"}]}'Attribute requests to your users
Set safety_identifier (or user) to your end customer's id. It becomes attribution_label in the request log and the usage export, filterable per customer. Retries, fallback and route pinning per request are on Model selection.
/api/v1/usageclient.chat.completions.create(
model="gpt-6-luna",
safety_identifier="customer-42",
messages=[{"role": "user", "content": "Hi"}],
)