What changes with Decisions API
OpenAI has made its Decisions API available to all customers, the model the company had announced during its developer day on September 29. TechCrunch reports this, according to which the product enters the market for fast, low-cost responses, the same market in which TypeSafe, a startup founded by a former researcher at a frontier lab, reached a valuation of $7.6 billion with the Jev model.
The category is that of models that act as super-classifiers: they receive a set of options and choose quickly, indicating the degree of confidence in the decision. According to the source, developers have begun using them instead of large language models in several workflows because they respond faster and cost much less. The new product is therefore OpenAI's response to a segment it did not create but now wants to establish a presence in.
Speed and multimodality as arguments
Nikunj Handa, head of product for OpenAI's APIs, told TechCrunch that the team took ChatGPT Luna and reconfigured it to reduce inference times and costs, without training a model specifically calibrated for decisions. The claimed improvement is in latency: from one to three seconds to 100, 120, and 180 milliseconds. Handa added that dedicated training could arrive in the future, but today the result is achieved with the existing model.
The public version also introduces voice and visual modes, and OpenAI has worked to lower latency across regions by geographically distributing inference resources. In early tests with customers such as Box, Datadog, Ramp, and Speak, engineers replaced OpenAI's language models with the new API to obtain greater processing capacity. Andrew Hsu, CTO of Speak, said the time between actions fell from 1.4 seconds to 0.13 seconds. OpenAI also used the product to make its own voice assistant Dots more responsive.
Benchmarks and caveats on evaluation
It is not clear what makes a decision model good: according to the source, the optimization among intelligence, cost, calibration, and speed varies depending on the workload. TypeSafe CEO Diogo Almeida is wary of benchmarks and encourages trying models on one's own work, even though TypeSafe published its own rankings alongside the latest version of Jev. The public indices AIM System One Index and Artificial Analysis Decision Index tend to favor Liquid's D1 Ultra and Base Lab's Elephant-1; Jev remains competitive, while OpenAI's model ranks eighth and fourteenth, respectively. TypeSafe remains the most queried on OpenRouter.
There is also JevMan, a benchmark in which models play Pac-Man: here Jev's latest model leads. According to Felix Wunderlich of Opper AI, who built the test, the game is interesting for its continuous stream of rapid, high-stakes decisions, with choices to be made in a few hundred milliseconds. Handa, by contrast, believes OpenAI's strength could be multimodality: for image classification Luna shows better calibration, and customers are asking for greater flexibility in using images as output. It remains to be verified whether the Luna-based approach holds up under real workloads; the source does not indicate prices, usage volumes, or availability times for additional features.
