OpenAI Decisions API pricing
GPT-6 Luna for fast application decisions: $0.10 per million input tokens, with no output, cache-read or cache-write charges.
Public beta announced 2026-10-06 · verified 2026-10-07
Official Decisions documentation ↗Decisions vs Responses: same model, different billing
The special rate applies to POST /v1/decisions using gpt-6-luna. Other Luna requests retain their model and processing-mode prices.
| USD per 1M tokens | Decisions API | Luna Standard Responses |
|---|---|---|
| Input | $0.10 | $0.10 |
| Output | No charge | $0.50 |
| Cache reads | No charge | $0.01 |
| Cache writes | No charge | $0.125 |
Base rates for prompts up to 272,000 input tokens. “No cache charge” does not mean all input is free. Decisions bills input tokens; the general calculator models ordinary generation. See Luna model pricing for the Responses rules.
Billing exceptions
Long-context input multipliers and regional processing premiums still apply. Luna's model documentation specifies 2× input rates for the entire request above 272,000 input tokens, plus a 10% regional-processing premium where available.
At base pricing, 100,000 requests with 1,000 billable input tokens each total 100 million tokens and cost $10. A 300,000-token request costs $0.06 with the long-context multiplier, or $0.066 with the regional premium as well. Image input must be counted in billable tokens; a photo is not a fixed per-image charge.
What it returns
Decisions accepts text and images. It supports three typed question formats:
- Predicates: probability that a condition is true.
- Choices: selection among predefined options, with confidence.
- Scores: evaluation against a numeric rubric.
Example applications include complaint routing, image checks, model selection and risk screening. Define the allowed outcomes before calling the endpoint.
How fast is it?
OpenAI reports up to 10× faster decisions than calling GPT-6 Luna through Responses. This is a task-specific vendor claim, not a guaranteed latency or a comparison with another provider.
First shown in limited preview at DevDay 2026, Decisions moved to public beta on October 6.
How it relates to Jev
TypeSafe AI's Jev also focuses on typed decisions, probabilities and confidence for software. We see overlapping product use cases. That does not establish equal accuracy or a speed advantage between the two; the 10× claim compares two OpenAI endpoints.
Usage tiers and token prices
OpenAI's new Build, Launch and Grow tiers qualify at $5, $100 and $500 in cumulative credit purchases. Higher calling capacity is separate from this endpoint's token price.