GPT-6 Luna Decisions is GPT-6 Luna powered by OpenAI's Decisions API.
Modality:
Text
Image
model ID
openai/gpt-6-luna-decisions
Output Speed *
113.24
tok/s
Intelligence Index *
27.9
/ 100
Context Window *
1050000
tokens
Input price
0.6
Anytoken
Output price
0
Anytoken
Performance
Bounded Answers Remove Generation From the Latency Path
GPT-6 Luna Decisions is optimized for bounded answers, not generation. OpenAI says the endpoint returns typed answers up to 10x faster than the same model through the Responses API. Sierra reports under 280 ms p95 when choosing the next action in live conversations. Both are provider-published claims, and OpenAI publishes no absolute latency or accuracy figures. Because the output is a probability, choice or score rather than prose, the decision step can sit inline in a request path. Measure your own p95 and calibrate thresholds on labeled data before gating production traffic.
Technical Specifications
What the model supports
The endpoint is POST /v1/decisions, and gpt-6-luna is the only model it currently supports. Inputs can be text, images or JSON. Outputs are always typed answers, never generated text. Two specifications matter most in production. The 400-question cap lets one shared input be evaluated against many questions in a single request. The underlying 1.05M-token window allows long shared evidence, though long-context input multipliers apply. Custom JSON schemas and tool calls are outside this endpoint and belong on the Responses API.
Verified Specifications —
GPT-6 Luna Decisions
*
Input modalities
Text
Image
output modalities
Text
Context window
1050000
tokens
Knowledge cutoff
September 2026
Pricing (standard)
0.6
AnyTokens in
/
0
AnyTokens out
Limitations & Trade-offs
1
No generated text or custom schemas. GPT-6 Luna Decisions returns probabilities, choices and scores, not prose. It doesn't accept custom JSON schemas, written explanations or tool-call requests. OpenAI directs those needs to Structured Outputs and function calling on the Responses API. That makes it unsuitable for extraction into your own fields, summarization, drafting or agent loops. Use it as a gate in front of generation, and use GPT-6 Luna on Responses or a larger model when the output itself is the product.
2
Luna-only ceiling and no published accuracy. Only gpt-6-luna is supported on the endpoint, so a hard judgment can't be escalated to a Sol or Astra tier inside Decisions. OpenAI publishes no accuracy or calibration figures, and the 10x speed claim carries no absolute latency. Subtle, high-stakes judgments such as legal, medical or financial calls need labeled evaluation sets, tuned thresholds and a fallback path, for example sending low-confidence cases to a stronger model on Responses.
3
Input-side cost and route-level caps. Billing covers input tokens only, so cost scales with the size of the shared input, not the answer. Long-context input multipliers and regional processing premiums apply, and no Batch or Flex tier is documented for this endpoint. Repeatedly sending large shared documents is the expensive pattern, and prompt caching helps with reused prefixes. Some gateways also enforce lower caps than OpenAI: one route lists a 66K-token context and another caps questions at 32. Verify limits on your route.
4
Thin documentation on context and effort. The reviewed OpenAI docs don't state a Decisions-specific context cap or any reasoning-effort control for this endpoint. Aggregators list context sizes from 66K to about 1.1M tokens, which reflects route policies rather than a single official figure. Teams planning long-document decisions or depending on deeper reasoning per request should test those cases directly and treat Luna on Responses, which documents effort levels, as the alternative.
Best-Fit Workloads
Where this model earns its place
01
Support ticket routing and triage
Choice questions route a message to billing or technical teams, and score questions rate severity against ordered levels. These are the patterns OpenAI uses in its own examples. Per-option probabilities and a confidence value let you set auto-route thresholds and send low-confidence tickets to a human or stronger model.
02
Real-time voice and agent orchestration
OpenAI added native Decisions integration in GPT-Live, so an app can act while a user is still speaking. In OpenAI's launch post, a customer reported cutting time to next action from 1.4s to 0.13s, and Sierra reported under 280 ms p95 for next-action selection. Both are customer-reported figures, so verify latency on your own traffic. Best for picking actions or reactions, not composing speech.
03
Image and structured-record checks
Predicate questions can check a product photo for visible damage, and per-question images let you compare candidates against a shared reference image. JSON input lets you test a ticket against an order record, such as a duplicate charge. Text, image and JSON inputs share one request shape. Accuracy figures for vision decisions are not published, so validate on labeled samples before automating refunds or moderation.
04
Rubric scoring and eval gating
Score questions grade an input against ordered levels and return a probability-weighted score, which suits rubric-style grading, prioritization queues and agent output checks. Up to 400 questions per request lets one shared input be graded on many criteria at once. OpenAI publishes no agreement figures against human graders, so calibrate against your own labeled set before using it as a release gate
Integration
01
One API integration
02
Easy model switching
03
Flexible for production
04