OpenAI’s Decisions API Bills Input Only, and Moves the Agent Off Text Generation

TL;DR

GA Decisions API runs on GPT-6 Luna for probabilities, choices, and rubric scores—not open-ended replies.
Uncached input: $0.10/M tokens; no output fee. A 10k-token context costs $0.001; 20M monthly decisions cost about $20,000.
Reported sub-300 ms p95 puts decisions in synchronous paths, including pre-action checks.
Verdicts route easily but offer no rationale; speed claims are company-reported.

A verdict, not a reply

OpenAI’s case is that agents make many small classification, routing, and next-action decisions whose latency and cost compound. The GA API focuses GPT-6 Luna on bounded questions over text, images, and structured inputs; OpenAI claims up to 10× the speed of Luna through Responses. It is less a chatbot than an interface: specify a narrow question, receive a typed result, let application code act. (Source: OpenAI — Build more responsive apps with the Decisions API)

The three answer types are distinct. A predicate estimates a condition’s probability from 0 to 1. A choice selects a predefined option and returns its probability distribution and confidence. A score rates input against ordered rubric levels, two to ten of them, and returns their probability-weighted average index, which can be fractional. These are structured judgments with uncertainty signals, not necessarily one label or scalar. GPT-6 Luna is currently the only supported model. OpenAI says dots uses Decisions to reduce response times and improve interactivity, Personal Finance to categorize transaction descriptions, and Communications to classify coverage by theme and score whether it reinforces or challenges narratives. (Source: OpenAI API Docs — Decisions; OpenAI — internal and customer examples)

Input becomes the billable surface

The headline rate is $0.10 per million uncached input tokens with no output-token charge, and that changes what is billable. A conventional generation bill grows with what a model reads and writes; Decisions returns a probability, supplied choice, or score distribution instead of open-ended prose. Luna’s generation endpoint was already priced at $0.10 in and $0.50 out per million tokens (September’s price-war post); Decisions keeps the input rate and drops output entirely. Inference is not costless, but the endpoint meters context rather than generated text. (Source: OpenAI — Decisions API announcement; OpenAI API Docs — Decisions pricing)

At that rate, 10,000 tokens cost 10,000 ÷ 1,000,000 × $0.10 = $0.001 per decision. At 20 million decisions monthly with that context size, the bill is about $20,000. The practical optimization target shifts from shortening an answer to not resending a bloated history: keep only state needed to decide and reuse stable input where caching applies. Cached rates differ; regional-processing premiums and long-context multipliers can also apply, so $0.10 is not every workload’s effective rate. (Source: OpenAI API Docs — Decisions pricing and caching; OpenRouter — GPT-6 Luna Decisions)

A model in the synchronous path

Sierra reports live next-action selection below 280 ms p95. Datadog scores the risk of agent-generated rshell commands before execution and reports 40× faster p95 with comparable accuracy. This puts the model in a remediation hot path, ahead of an automated action. (Source: OpenAI — customer examples)

At that speed, a decision model becomes a synchronous service dependency. Its timeout, availability, thresholds, and fallback belong in the action path’s reliability design. The model is no longer just a generator the agent consults; it gates what the agent may do. Measure false blocks and missed risks alongside latency. (Source: OpenAI — customer examples)

A score is not a reason

Bounded answers are easy to parse and threshold, but do not explain themselves. Predicates give a probability; choices and scores include probability information and confidence. None is a rationale. A Datadog risk score can be audited as a number, not as the reasoning behind a command’s classification.

For autonomous remediation, that changes the evidence available after an action. OpenAI advises calibrating thresholds on labeled examples against the costs of false positives and negatives. Preserve the input, decision version, returned probabilities, threshold, and resulting action; add review or a separate explanation where policy requires a rationale. (Source: OpenAI API Docs — Decisions and interpretation)

Four hundred questions change the unit

The GA limit is 400 questions per request, twice the previous cap. Independent questions can share input; dependent ones require separate requests. The release also adds native GPT-Live integration, caching, US and EU inference residency, native JSON, images in question instructions, and capacity in several Asian regions. (Source: OpenAI — Decisions API announcement)

Batching makes a request the shared-context unit, while each question remains a quality unit. Teams can evaluate several independent properties of a document together. Report accuracy, calibration, and error costs per question, plus latency and spend per request; otherwise a healthy aggregate can hide a weak question. Dependencies remain sequential. (Source: OpenAI API Docs — Ask multiple questions)

Treat customer numbers as signals

Box reports “~5x faster route selection on 30 certificate of insurance cases, with the same accuracy as the Responses API.” Ramp says it was “4x faster and 30% cheaper to attribute our AI costs to the actual work when compared to GPT-6 Luna, without sacrificing accuracy.” Datadog reports “40x faster p95 with comparable accuracy” against available frontier models, scoring rshell risk before execution. Sierra reports “under 280 ms p95, about 3× faster and 90% lower cost per call than our previous approach, with no drop in customer outcomes.” Speak says it could “reduce the time to next action by over 90%, from 1.4s to 0.13s.” The company-reported evaluations use different tasks and baselines; they are not independent or comparable tests, and the release includes no third-party benchmark. That is an absence of evidence, not proof the claims are false. (Source: OpenAI — customer examples)

The signal is practical, not universal: teams are testing bounded decisions for routing, allocation, orchestration, and pre-action risk. It shifts repeated judgments into software; builders must keep context lean and accept that a verdict is not its own explanation.

FAQ

Is output free?

There is no output-token fee. Input rates vary with caching, region, and context length.

Does it return one label?

No. Predicates return probabilities; choices and scores include probability data and confidence.

Can questions depend on each other in one call?

No. Batch independent questions; dependent questions require separate requests.

Are customer speed claims independently verified?

The cited release gives customer evaluations, not an independent benchmark.

OpenAI's Decisions API Bills Input Only, and Moves the Agent Off Text Generation