GPT-6 Luna Decisions on Lazu: typed answers via API
Lazu supports the GPT-6 Luna Decisions endpoint. Use /v1/decisions to obtain typed answers for predicates, fixed choices and rubric scores, such as whether a support ticket needs urgent attention.
Start with a task you can verify #
The catalog and key permissions use gpt-6-luna-decisions; the official request model is gpt-6-luna. This route has separate pricing from Chat and Responses. Use question names to match answers to application fields, then apply your own business rules.
Pricing and boundaries #
This is a snapshot of Lazu’s stable-lane base rates on October 10, 2026, before any account-specific discount. Consult the catalog for subsequent changes.
| Dimension | USD / million tokens |
|---|---|
| Input | $0.12 |
| Output | $0 |
| Cache read | $0 |
Lazu currently limits input plus questions to 8192 estimated text tokens, 64 questions and four inline images. Streaming and tools are unsupported. A per-question refusal can arrive in HTTP 200 and remains a billed result. Text and image input cost $0.12 per million tokens; output is currently priced at zero.
The model card below reads the live catalog:
gpt-6-luna-decisionsOpenAI · 8k · Stable $0.12 / $0.00Model page →A minimal request #
Create a key with a model restriction and spending cap, then set LAZU_API_KEY in your local environment. Keep the key out of source code. Query /api/models/catalog with that key to check permissions and supported endpoints before integration.
curl --fail-with-body https://api.lazu.ai/v1/decisions \
-H "Authorization: Bearer $LAZU_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"gpt-6-luna","input":"The parcel arrived damaged.","questions":[{"name":"damaged","type":"predicate","instructions":"Did the parcel arrive damaged?"}]}'
Verify the result before increasing traffic #
On October 10, 2026, a restricted key completed this short-text request through the public API with HTTP 200 and a valid response. This verifies one basic request path, not sustained concurrency, long-context behavior or task quality.
Save X-Lazu-Request-Id from the response headers. Inspect the input, output, cache usage and charge in Usage & Logs, or query GET /api/usage/requests/{request_id} with the same key. Distinguish quota and budget errors from upstream rate limits; retry only when appropriate and with a limit.
See the Lazu catalog, pricing and lanes, and model reference.