OpenAI has made its Decisions API model publicly available to all users, a move that sees the developer enter the market for fast, cheap AI, where TypeSafe, a startup recently founded by a former OpenAI researcher, achieved a $7.6 billion valuation thanks to its Jev model.
OpenAI has made its Decisions API model publicly available to all users, a move that sees the developer enter the market for fast, cheap AI, where TypeSafe, a startup recently founded by a former OpenAI researcher, achieved a $7.6 billion valuation thanks to its Jev model.
Nikunj Handa, OpenAI’s head of APIs, told TechCrunch that after Jev’s success, his team quickly developed their own alternative. Rather than building a specially calibrated model from scratch, engineers reconfigured ChatGPT Luna to speed up data output and reduce the cost of use.
“The main achievement of the Decisions API is speed: response time has decreased from 1–3 seconds to 100–180 milliseconds,” Handa noted. “In the future, we may conduct special training of models for such tasks, but for now, regular Luna copes with this perfectly.”
Announced at the company's developer conference on September 29, the Decisions API became available to the general public on October 22. OpenAI added voice and visual modalities to the public version, and also optimized the allocation of computing resources to reduce latency around the world.
In previous trials at Box, Datadog, Ramp, and Speak, engineers have used the new API to replace standard large language models to increase throughput. Andrew Hsu, CTO at Speak, said using the Decisions API reduced the time between actions in their app from 1.4 seconds to 0.13 seconds. OpenAI has also used the new solution to speed up its voice assistant Dots.
There are currently no clear standards for decision models, as optimizations for accuracy, cost, and speed depend on the specific task. TypeSafe co-founder Diogo Almeida encourages testing models directly on your own workloads, although his company recently released its own ratings along with a new version of Jev that added support for images and conditionals.
According to public tests (AIM System One Index and Artificial Analysis Decision Index), the leadership is held by the D1 Ultra model from Liquid and Elephant-1 from Base Lab. The Jev model occupies 7th and 3rd places in these lists, respectively, while the OpenAI solution is located in 8th and 14th places. Despite this, the TypeSafe models currently remain the most popular in terms of the number of requests on the OpenRouter platform.
Of particular note is Opper AI’s JevMan entertainment benchmark, where models compete in the game PacMan. TypeSafe’s development holds the lead in it. According to Opper AI representative Felix Wunderlich, the classic game is an excellent test of quick decision-making, as the model has to choose a direction of movement in a matter of hundreds of milliseconds.
The main advantage of OpenAI's solution may be multimodality. For image classification, the Luna model shows high accuracy, and users have already expressed interest in using the Decisions API for generating and processing visual content.
