TL;DR
- General Release: OpenAI’s Decisions API is generally available, giving developers probabilities, choices and scores through its GPT-6 Luna model.
- Voice Apps: Integration with OpenAI’s GPT-Live voice service sends conversation transcripts directly for decisions, removing an application round trip.
- Input Costs: New prompt caching charges to store shared material and discounts the input price when that material is reused.
OpenAI has moved its Decisions API, which gives software probabilities, choices or scores instead of written replies, into general availability on October 22. For applications making repeated small judgments, OpenAI claims about tenfold faster answers than its text-generating Responses API. Andrew Hsu, chief technology officer at AI language-tutoring company Speak, told TechCrunch that its trial cut the interval between actions from 1.4 seconds to 0.13 seconds.
Developers already had access through an all-developer public beta on October 6. The general release adds prompt caching and native voice integration, while doubling the maximum number of questions in one request from 200 to 400.
Small Judgments Instead of Written Replies
Decisions API runs on GPT-6 Luna, the model also available through OpenAI’s Responses API. Nikunj Handa, OpenAI’s API product lead, told TechCrunch that TypeSafe’s rival Jev decision model showed his team uses for its own version. His team reconfigured Luna for faster, cheaper inference rather than training a new model.
The Decisions API evaluates text, images or structured JSON data and returns a probability, one of the developer’s supplied choices, or a score against ordered criteria. Responses can generate a custom data object, a written explanation or a request to use a tool with specified arguments.
As OpenAI demonstrates with a product-photo example, an application can ask whether the image contains a crack, tear or dent. The Decisions API then returns an estimated probability of damage, and the application uses a chosen cutoff to flag the photo for review. Developers can set thresholds using labeled examples and the relative costs of false alarms and missed cases.
One request can ask independent questions about the same photo, such as whether it is damaged and which product category it belongs to. A later question that depends on the first answer still needs another request. The application performs any action selected by the answers.
A Shorter Route for Voice Apps
OpenAI’s GPT-Live voice service manages the spoken conversation. Its native connection to the Decisions API sends conversation transcripts directly for judgments while a user is still speaking. With custom delegation, GPT-Live notifies the application, which then calls the Decisions API; the native connection removes that round trip.
In OpenAI’s documented car-assistant example, a driver saying they feel cold lead Decisions API to select a temperature increase, which the application carries out. A request for directions and restaurants can be handed to Responses and its tools to do the search.
Image Support and Deployment Differ
Both OpenAI’s generally available service and TypeSafe’s Jev support image judgments. TypeSafe labels its image features experimental and says it cannot vouch for them in production.
Liquid AI’s d1-ultra is another hosted option that reads text and images and returns probabilities, choices or scores for decision making.
Base Labs, Baseten’s research team, released elephant-1 and mammoth-1 with downloadable model weights. Developers can use those model files to run the models on their own systems. A one-week free trial is also available on Baseten’s servers.
Classification Strength Can Diverge From Browser Success
In AIMultiple’s independent decision-model tests, classification leader decider-12b completed 16 of 50 browser tasks. Kev-27B, which ranked below Jev on classification, completed 40. Assigning the right label to a record and choosing a successful sequence of browser actions tested different abilities.
In their rests, the browser setup turns each page into text and a list of controls. Models select an operation and target; the software executes the action and reads the page again. A separate text-generating model supplies words for form fields, because the decision models returns option scores rather than text.
Caching Changes the Input Bill
At standard rates, OpenAI’s Decisions API costs $0.10 per million uncached input tokens, the same Luna input rate as for the Responses API. Decisions API has no output-token charge; standard Luna output through Responses API costs $0.50 per million tokens.
In an illustration from consultancy BaristaLabs, a million calls with 2,000 input tokens each cost $200 through Decisions, versus $225 through Responses if each answer contains 50 output tokens. The illustration excludes caching.
By default, repeated requests can reuse an identical prefix of shared input once it reaches 1,024 tokens. A reference document or image can stay fixed while the questions change; question content falls outside the reusable prefix. Under Luna’s standard caching rates, reads cost $0.01 per million tokens, while writing a cache carries a charge. Requests configured to keep processing in a specified region can carry an extra charge.
Liquid lists d1-ultra at $0.042 per million input tokens, but bills each question as its own prompt, including its text and images.