Introduction
GPT 5.6 Luna is designed for cost-sensitive work at high volume. It fills the small-model role in the GPT 5.6 family, accepting text or images and returning text. Its 1,050,000-token context can carry common reference material, but a queue of independent records usually benefits from keeping each request focused on the item being processed.
Key features & highlights
Luna supports structured output for labels, extracted fields and compact transformations that another program will consume. A constrained result is easier to validate and retry when a record is incomplete. A successful parse still needs content checks: an allowed category or a correctly typed amount can be wrong for the source.
Reasoning can be disabled or set to low, medium, high, xhigh or max, with medium as the default. This gives a queue a way to spend more effort on difficult records without changing its result structure. Function calls can retrieve application-owned details when needed, while the application decides whether the requested operation is permitted.
Request parameters
| Parameter | Required | Description |
|---|---|---|
messages | Yes | Design one stable contract for the queue and include only examples that teach a difficult distinction. at least 1 items · at most 256 items |
max_tokens | No | Product coding — The output limit covers reasoning and visible text together. |
reasoning_effort | No | Product coding — Select none, low, medium, high, xhigh or max, The default is medium. |
temperature | No | Product coding — Temperature applies only at reasoning effort none. |
top_p | No | Product coding — Use |
tools | No | Product coding — Describe the functions your application can execute and validate their arguments before running them. at most 128 items |
parallel_tool_calls | No | Product coding — Independent function requests can arrive together, The application controls their execution. Default: |
tool_choice | No | Product coding — Allow automatic selection, forbid calls, require a call or pick one named function. |
response_format | No | Product coding — Choose plain text, a JSON object or a JSON Schema response. |
Pricing
Input $0.16 / Output $0.96 per 1 million tokens; Prompts over 272K tokens: Input $0.32 / Output $1.44
Use cases
For a support intake queue, define the categories, urgent conditions and an explicit unknown outcome. Include a few boundary cases in the instructions. Route unclear messages to review rather than allowing a mandatory label to conceal missing evidence.
For invoice preparation, provide a readable document image and ask Luna to preserve the invoice identifier, currency and line amounts in separate fields. Compare totals and source identifiers in code before writing a record to the accounting system.
Prompt guide & pro tips
Design one stable contract for the queue and include only examples that teach a difficult distinction. State which source field supplies each output and how absent information should be represented. Keep identifiers as strings when leading zeroes matter. For a rewrite, specify the maximum length and the facts that must survive unchanged.
Begin with a representative labeled evaluation set and measure accepted results, retries and review volume together. A smaller output cap limits generation but includes reasoning, so an exhausted cap may leave no usable answer. Use reasoning none for explicit temperature or top_p adjustments, and inspect the terminal status before accepting an item. Separate independent requests with bounded concurrency in your own worker; do not treat a model's high-volume positioning as a guarantee of a particular rate limit or response time.
API quick start
curl --fail-with-body -sN https://api.spicyapi.ai/v1/chat/completions \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"model": "openai/gpt-5.6-luna/chat",
"messages": [
{
"role": "user",
"content": "Summarise the following release notes into three bullet points."
}
],
"stream": true
}' \
| jq --unbuffered -Rrj 'select(startswith("data: ")) | .[6:] | select(. != "[DONE]") | fromjson | if .error then "\n[error] \(.error.message)\n" else (.choices[0].delta.content? // empty) end'
See the API reference for authentication, the complete schema, polling, and error handling.

















