General availability just over 2 weeks after the trial version
OpenAI on Oct. 22 (Oct. 23 Japan time) opened the "Decisions API," which returns only the answer to a choice or classification, to all developers. It had released the API as a trial version on Oct. 7. @OpenAIDevs, the company's X account for developers, made the announcement.
Given a question and options, the Decisions API returns only an answer and probabilities, without writing text. It comes in three forms: one returns the probability that a condition is met, one picks a single option from candidates, and one returns a graded rating. Inputs can include images and JSON as well as text. It is intended for uses such as routing inquiries, judging posts and deciding an agent's next action.
Responses in about 0.1 seconds
According to Nikunj Handa, who leads the effort, response time fell from 1 to 3 seconds to 0.1 to 0.18 seconds. The company says it is up to 10 times faster than having its lightweight model "GPT-6 Luna" do the same task through the regular API. It did not build a new model but changed how Luna is run. U.S. language-learning app Speak said its processing time fell from 1.4 seconds to 0.13 seconds.
With general availability, the number of questions that can be sent in one request was doubled from the trial version to 400. The API can now be connected to "GPT-Live," the voice conversation feature, and users can choose to have processing done in the United States or Europe. Box, Datadog and Ramp of the U.S. have also been early users.
Price about 2.4 times that of a specialist
Pricing is $0.10 per 1 million input tokens, with no charge for output. Reusing the same prompt prefix lowers the input price to $0.01. "Jev" from U.S.-based TypeSafe, which specializes in the same use, charges $0.042 for input, and OpenAI's regular input price is about 2.4 times that.
| Item | OpenAI Decisions API | TypeSafe Jev |
|---|---|---|
| Input (1 million tokens) | $0.10 | $0.042 |
| Reused input | $0.01 | No separate price |
| Output | No charge | No charge |
Jev ranks higher in evaluation
In a judgment evaluation by third-party U.S. firm Artificial Analysis, Jev ranked third and OpenAI's model ranked 14th. Both companies' speed figures are their own measurements and cannot be compared directly. Handa said that if responses reach 0.06 seconds, it will become possible to make a judgment on every frame of a video.
