# OpenAI Decisions API Drops Output Fees With GPT-6 Luna

URL: https://3ammarketer.com/news/openai-decisions-api-pricing-benchmarks
Published: 2026-10-23

> OpenAI made the Decisions API generally available on October 22, 2026, cutting output charges completely and evaluating up to 400 questions per request on GPT-6 Luna.

OpenAI made the Decisions API generally available on October 22, 2026, eliminating output-token fees entirely by charging only $0.10 per million input tokens. Built with GPT-6 Luna, the dedicated endpoint processes text, images, and structured data up to 10x faster than GPT-6 Luna through the standard Responses API by evaluating bounded questions directly. [1][2][3]

## What is the Decisions API and how does it evaluate inputs?

The dedicated HTTP interface is exposed at the POST /v1/decisions endpoint running on GPT-6 Luna. Instead of streaming text token by token through autoregressive generation, the endpoint focuses model intelligence specifically on bounded questions across text, images, or native JSON inputs. Because the workflow produces structured evaluation objects rather than generative prose, OpenAI records output_tokens as 0 in usage payloads and charges zero output token fees. [1][2]

## The four answer types: predicate, choice, score, and refusal

The API answers questions across three primary configured types alongside a fallback safety state. A predicate question calculates the probability that an evaluated condition is true, outputting a value from 0 to 1. A choice question selects one option from 2 to 255 unique user-supplied choices, returning the selected choice together with explicit probabilities and confidence. A score question grades the input against ordered levels within a rubric, calculating a probability-weighted average of the level indices. [1][2]

A fourth format, refusal, can return for any individual question in the response array rather than failing the entire batch request. This modular error reporting preserves valid answers for adjacent questions when one item triggers a boundary refusal. [2]

## Input-only pricing, cache discounts, and regional surcharges

Billing operates solely on input volume at $0.10 per million input tokens, dropping to $0.01 per million tokens for cached input. On general endpoints, standard GPT-6 Luna access charges $0.50 per 1M output tokens alongside that same $0.10 input rate, meaning the Decisions API strips out generation expenses entirely. According to our [LLM API pricing timeline](/data/model-pricing), running 1 million input plus 1 million output tokens on standard GPT-6 Luna totals $0.60, whereas the specialised GPT-5.6-cyber reaches $87.50 for the same volume. [1][3][4]

![OpenAI business pricing table listing GPT-6 Luna standard input, cached input, and output token rates.](https://dygolixznokmekyvajbb.supabase.co/storage/v1/object/public/images/content/17f5fa07272b407f3fe3c067.png)

Comparison of combined cost for 1 million input and 1 million output tokens across select OpenAI models in October 2026. [4]

- GPT-6 Luna: 0.6 $
- GPT-5.6-cyber: 87.5 $

Under this structure, a decision containing 400 input tokens costs $0.00004, which equals $40 per million decisions. However, two pricing adjustments apply to specialized traffic: inputs exceeding 272K tokens are billed at a 2x rate ($0.20 per million tokens), and routing requests through dedicated US or EU data-residency endpoints adds a 10% premium. [2][5]

## What early production benchmarks show

Early production migrations show marked latency gains: Box reported ~5x faster route selection on 30 certificate of insurance cases with equal accuracy to the Responses API, Datadog recorded 40x faster p95 latency scoring agent command risk, and OpenAI reported that integrating into GPT-Live reduced action time by over 90%, from 1.4s to 0.13s. [1]

Speak GPT-Live action time before and after integrating the Decisions API in October 2026. [1]

- Before Decisions API: 1.4 s
- With Decisions API: 0.13 s

Customer evaluation also showed the endpoint runs under 280 ms p95, approximately 3x faster and at 90% lower cost per call than previous approaches with no drop in customer outcomes, while cost-attribution workflows ran 4x faster and 30% cheaper compared to GPT-6 Luna without sacrificing accuracy, proving effective for high-frequency routing across [autonomous AI agents](/blog/autonomous-ai-agents-runtimes-costs). [1]

## Payload anatomy: questions, inputs, and safety identifiers

A valid POST /v1/decisions request body contains model (set to gpt-6-luna), input (a plain string or an array mixing input_text and input_image elements), and questions. Developers can supply up to 400 questions per request, doubling the prior beta ceiling. Callers can also include an optional safety_identifier parameter carrying an end-user identifier up to 128 characters long. [1][2]

Limit: The Decisions API does not generate freeform conversational text, markdown, or unstructured explanations. Applications that require natural-language rationale alongside classifications must query generative endpoints instead.

Engineers routing classification, routing, or guardrail checks through general chat endpoints should move those workloads to POST /v1/decisions to remove output billing and cut downstream queue delays. For workflows needing unstructured responses, standard Responses endpoints remain mandatory. [1][2]

## Sources

1. [Build more responsive apps with the Decisions API | OpenAI](https://openai.com/index/decisions-api)
2. [What Is the OpenAI Decisions API?](https://apidog.com/blog/openai-decisions-api/)
3. [Business Pricing | OpenAI](https://openai.com/api/pricing/)
4. [LLM API pricing timeline (3AM Marketer data set)](https://3ammarketer.com/data/model-pricing)
5. [OpenAI Decisions API Pricing: $0.10/1M Input, Cost per Decision - Public Beta on GPT-6 Luna, How to Call It, and Limits | AI Free API](https://www.aifreeapi.com/en/posts/openai-decisions-api)