OpenAI wants developers to stop paying for paragraphs when all they need is a verdict. On October 22 the company made its Decisions API generally available, a dedicated endpoint that asks GPT-6 Luna questions with a fixed set of possible answers and returns a structured result rather than prose. OpenAI first showed the product at its DevDay event on September 29 and had run it as a public beta since.
The design rests on three answer formats. A predicate question returns the probability, from 0 to 1, that a stated condition holds, for example whether a photo shows a damaged parcel. A choice question picks one option from a list the developer supplies, such as continue, retry or ask the user, and reports probabilities and a confidence value alongside it. A score question places the input on a ladder of ranked rubric levels and returns a probability-weighted position. Context can be passed as text, images or native JSON.
Several additions arrive with general availability. OpenAI lists a native link to GPT-Live, its real-time voice stack, so a voice agent can decide mid-conversation whether to clarify, route or escalate; prompt caching for repeated context; processing kept in the US or the EU; images inside question instructions; and a cap of 400 questions per request, double the beta limit. The company also says it added capacity in several Asian regions to cut latency.
Pricing is the most unusual part. Developers pay $0.10 per million input tokens, $0.01 per million for cached input, and nothing for output. OpenAI says the endpoint answers up to ten times faster than putting the same question to GPT-6 Luna through the Responses API. Customer figures in the announcement are self-reported: Datadog cites a 40-fold improvement in p95 latency when scoring the risk of agent-generated shell commands, Sierra reports calls under 280 milliseconds at the 95th percentile and about 90 percent lower cost per call, Ramp says cost attribution became four times faster and 30 percent cheaper, and Speak says the gap before its tutor's next action fell from 1.4 seconds to 0.13.
The launch places OpenAI in a young market for so-called decision models. TechCrunch reports that OpenAI moved after seeing TypeSafe's Jev, a model built specifically for fast classification with confidence ratings. Nikunj Handa, who leads product for OpenAI's APIs, told the publication that the company did not train a separate model but reconfigured Luna for faster, cheaper inference, cutting responses from one to three seconds to roughly 100 to 180 milliseconds.
Independent rankings do not yet favor OpenAI. On two third-party leaderboards cited by TechCrunch, the Artificial Analysis Decision Index and the AIM System One Index, OpenAI's entry sits fourteenth and eighth respectively, behind Liquid's D1 Ultra, Base Labs' Elephant-1 and TypeSafe's latest Jev, which also draws the most queries on OpenRouter. Handa argues that Luna's image handling and calibration are where OpenAI can compete.
For teams building agents, the appeal is placing a cheap, bounded check inside the hot path of a workflow, such as gating a tool call or routing a ticket, without paying for long generations. Whether the latency and accuracy figures hold outside the quoted pilots, and how well the calibration of a general-purpose model compares with purpose-built rivals, is something developers will have to measure on their own workloads.