OpenAI has opened its OpenAI Decisions API to every customer, moving directly into a market that TypeSafe built with its Jev family of decision models. General availability arrived on Thursday, following an announcement at OpenAI’s developer day on September 29.
What the API actually is
The Decisions API is not a purpose-built decision model in the way Jev is. OpenAI built it by reconfiguring ChatGPT Luna, with inference speed and cost as the design targets. The public release adds voice and visual modalities, and OpenAI has begun directing inference resources geographically to reduce latency. No specific calibration work has been described for the current version.
Latency is the headline
Nikunj Handa, product lead for OpenAI APIs, said latency has fallen from a range of one to three seconds to between 100 and 180 milliseconds. He described the speed improvement as the main unlock. His longer-term target is 60 milliseconds, which he said could support per-frame video API calls.
Speak, one of the early trial customers, reported that action time dropped from 1.4 seconds to 0.13 seconds. Andrew Hsu, CTO at Speak, confirmed the reduction. OpenAI also used the API to sharpen the responsiveness of its Dots voice assistant. Box, Datadog and Ramp joined Speak in the initial trials.
Why decision models exist
Decision models are fast, inexpensive alternatives to large language models for classification-style work. They return a choice with a confidence rating, which makes them suitable for workflows where a full LLM is slower and costlier than necessary. TypeSafe, founded by a former OpenAI researcher, created the category with Jev and is valued at $7.6 billion. Handa said colleagues saw many use cases for Jev, which is how OpenAI arrived at the Decisions API.
Benchmarks, and the fight over them
There is no settled definition of a good decision model beyond whether it works for a given workload. The right balance of intelligence, cost, calibration and speed shifts from one task to the next. TypeSafe’s latest Jev model, released last week, supports images and smart-if statements. Liquid released D1 Ultra in the same week.
On the AIM System One Index and the Artificial Analysis Decision Index, D1 Ultra and Elephant-1 lead. Jev ranks seventh and third on the two indices, while OpenAI’s decision model ranks eighth and fourteenth. TypeSafe also published its own rankings alongside the new model. Diogo Almeida, CEO of TypeSafe, distrusts benchmarks and urges customers to run their own trials.
Query volume on OpenRouter currently favors TypeSafe’s offerings. The JevMan benchmark, which has decision models play PacMan, is topped by the latest Jev model. Felix Wunderlich, chief growth officer at Opper AI, said PacMan tests rapid, high-stakes decisions.
Multi-modality and what comes next
Handa said Luna is powerful for classification and offers better calibration for that use. Customers have asked for flexibility in outputs. He also said future specific training may improve decision models beyond the current plain-Luna configuration, though details of that training have not been disclosed.
