OPENAI MODEL GUIDE
GPT-5.6 Luna API
The cost-sensitive GPT-5.6 tier for high-volume workloads, fast routing, extraction, and lightweight reasoning.
Classification, extraction, content operations, narrow sub-agents, and cost-sensitive high-volume traffic.
Unsupervised complex agents, difficult repository changes, and high-risk decisions that cannot tolerate retries.
Read by decision
What Luna should handle
Luna fits high-volume requests with narrow instructions and machine-checkable results. Its low rate becomes real savings only when output is stable, retries are bounded, and failures can escalate.
- Classification, extraction, moderation routing, and data transformation.
- High-volume support drafts and content operations.
- Sub-agents with narrow instructions and explicit validation.
- Cost-sensitive applications that still need a large context ceiling.
Design a Luna → Terra → Sol routing waterfall
The application should own escalation rules explicitly and record which model ultimately completed the task.
| Tier | Work handled | Escalation signal |
|---|---|---|
| Luna | Classify, extract, transform, draft | Schema failure, low confidence, retry limit |
| Terra | General coding, tools, complex inputs | Quality gate fails or sample is high risk |
| Sol | Hardest and highest-value work | Human review or stop |
A low token rate is not a low task cost
- Vendor list price is $0.20 per million input tokens and $1.20 per million output tokens.
- Cached input is published at $0.02 per million tokens.
- For high-volume workloads, track cost per accepted result because retries can erase the apparent rate advantage.
Build a machine-checkable output contract first
- Use structured outputs or a strict JSON schema.
- Validate missing fields, invalid enums, and empty results.
- Bound automatic retries and separate retryable from terminal errors.
- Escalate low-confidence or high-risk samples to Terra.
Minimal Luna request
Validate the route, output format, and usage record with a small non-streaming request before increasing concurrency or adding tools.
curl https://app.llmfly.ai/v1/chat/completions \
-H "Authorization: Bearer $LLMFLY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-5.6-luna",
"messages": [{"role": "user", "content": "Summarize the tradeoffs in three bullets."}],
"stream": false
}'What to test before high-volume rollout
- Measure pass rate on the real input distribution, not ideal examples.
- Track P50/P95 latency, 429s, retries, and escalation share together.
- Check for accidental long-context use; a large ceiling does not require resending all history.
- Alert on cost and error-rate anomalies so retry storms do not amplify spend.
Luna capability boundaries
- Lower token price does not prove suitability for complex autonomous work.
- Long context can still increase latency and spend if prompts are not curated.
- Evaluate schema adherence and tool recovery on real failure cases.
- Use Terra or Sol as fallbacks when quality thresholds are not met.
GPT-5.6 Luna specification sources and usage notes
Specifications and vendor prices were checked against OpenAI model documentation: https://developers.openai.com/api/docs/models. Confirm the current model ID, billing multiplier, and availability in Model Plaza.
Frequently asked questions
Why is Luna suited to high-volume work?
Its vendor reference rate is low while it retains structured outputs, tools, and a large context ceiling, but real pass rate and retry cost still require validation.
What should happen when Luna fails?
Use bounded retries and an explicit escalation rule, usually to Terra first and Sol only for high-risk or hardest samples.
Can Luna run complex agents?
Test it for narrow sub-agents with validation, but do not choose it for unsupervised complex autonomy based on price alone.
What is Luna's context window?
The vendor publishes 1.05M context and 128K maximum output, though large prompts still add latency and cost.
How do I know whether Luna really saves money?
Track total cost per accepted result, including retries, escalation, tool loops, and human rework.
Test GPT-5.6 Luna with your own task
Create an API key, confirm the model ID, and start with a small representative request.