API reference

Two ingress endpoints, one receipt shape, and an honest list of what is not built yet.

Authentication

Every request carries an API key as a Bearer token. Anthropic clients may send it as x-api-key instead, because that is what their SDKs do.

header

Authorization: Bearer llm11_live_...

Keys prefixed llm11_sandbox_ behave identically and hit real models, so their receipts are real. They are flagged as sandbox on every request, which is what lets you keep test traffic out of the figures you report.

POST /v1/chat/completions

OpenAI-compatible. Accepts messages, temperature, max_tokens and response_format, plus three additions.

request

{
  "messages": [{ "role": "user", "content": "…" }],
  "context": "optional source text the groundedness check holds claims against",
  "model": "llm11-balanced",
  "llm11_models": ["openai/gpt-5.6-luna", "anthropic/claude-opus-5"],
  "response_format": { "type": "json_object" }
}

model is a routing pack (llm11-fast, llm11-balanced, llm11-smart, or llm11-auto for the project default), or a specific catalogue model id, which pins the request with no routing at all. llm11_models is an explicit candidate pool that overrides both model and the project default; an id it does not recognise returns a 400 rather than silently dropping it. Sending response_format as json_object also turns the schema check from a shape test into a parse test, so malformed JSON escalates rather than reaching your parser.

POST /v1/messages

Anthropic Messages-compatible. Same engine, same receipt, translated at the edges. Point an Anthropic SDK at this origin and it works unmodified.

The receipt

Returned as _llm11 in the response body, with the headline figures mirrored into headers.

Field	Meaning
verdict	pass, escalated or failed.
modelUsed	The model that produced the answer you received.
baselineModel	The most expensive model in the offered pool, used as the savings denominator.
candidatePool	Every model id triage was actually offered for this request.
poolSource	request_list, request_pack, project, pinned or fallback: where the pool came from.
taskType	What triage decided this request was.
criticality	low, normal or high, after blending your policy floor with the triage call.
triageBackend	jev or heuristic. Named on every request rather than assumed.
triageConfidence	Calibrated where the backend is Jev, a coarse estimate otherwise.
rungsRun	Each check that ran, whether it passed, and its confidence where it has one.
escalated	Whether a failed check sent the request to the pool's strongest remaining model.
escalationReason	Which check failed, in words.
costCents	Generating the answer: the primary call plus any escalation.
verificationCostCents	Checking the answer. Never overlaps costCents.
triageCostCents	The routing decision itself. Real for Jev, genuinely 0 on the heuristic backend.
triageCostEstimated	True when triageCostCents is estimated from token count rather than reported.
totalCostCents	The three above, summed. What this request debited.
baselineCostCents	The same tokens priced on baselineModel, or null if it could not be priced.
savingsCents	baselineCostCents minus totalCostCents, or null if the baseline was null.
cacheLossCents	Always null today. Provider cache telemetry is not collected yet, and we will not estimate it.
latencyMs	End to end, including every check.
triageLatencyMs	The triage call alone.

costCents, verificationCostCents and triageCostCents are disjoint and sum to totalCostCents, which is what the request debits. Anything we could not compute comes back null, never zero, so a missing number is never mistaken for a good one.

response headers

x-llm11-request-id: 7c2f…
x-llm11-model: openai/gpt-5.6-luna
x-llm11-verdict: pass
x-llm11-escalated: false
x-llm11-cost-cents: 0.0041
x-llm11-baseline-cost-cents: 0.0462
x-llm11-savings-cents: 0.0412

Errors

Structured, with a stable type you can branch on.

error

{ "error": { "message": "…", "type": "insufficient_credits" } }

401	invalid_request_error	Missing, malformed or revoked API key.
400	invalid_request_error	Body was not JSON, or messages was absent or empty.
400	invalid_pool	model or llm11_models named a pack or id that does not exist in the live catalogue.
402	insufficient_credits	Prepaid balance is zero. Top up to continue.
429	rate_limited	Too many demo requests from one address. Applies to the public homepage demo only.
501	not_implemented	Streaming was requested and is not built yet.
502	upstream_error	The model provider failed. The message carries their reason.

Limits, stated plainly

There is no volume cap and no feature tier. Every request, every model, every check and every triage call runs at provider list price out of your prepaid balance, whatever that balance is. At zero, requests stop with a 402 rather than running the balance negative or quietly turning into a surprise invoice.

Streaming is not implemented. stream: true returns a 501 rather than a non-streaming response dressed up as one. Verification needs a finished answer, so streaming needs a design that holds the stream until the checks clear, and shipping it half-done would mean either lying about the format or streaming unverified tokens.

Machine-readable

/api/v1/openapi.json for the full schema.
/llms.txt for a summary written for an assistant to read.
/pricing.json for the credit packs and purchase fee as data.