OpenAI has made a series of moves spanning developer tools, enterprise adoption and AI safety, opening its Decisions API to all customers, expanding its partnership with Uber and moving frontier model training away from the live internet.
The Decisions API, first announced at the company’s September 29 developer day, is now generally available with added voice and visual capabilities and lower latency through geographically targeted inference. It enters a market for fast, low-cost decision models created by TypeSafe, a startup founded by a former OpenAI researcher and now valued at $7.6 billion. TypeSafe’s Jev model pioneered a category of super classifiers that choose between options and give confidence ratings, often replacing large language models at far lower cost and higher speed.
Nikunj Handa, product lead for OpenAI’s APIs, said his colleagues saw countless use cases for Jev, so the team reconfigured ChatGPT Luna for faster, cheaper inference rather than building a dedicated model. He said the key advance is speed, cutting response times from one to three seconds to between 100 and 180 milliseconds, and suggested future gains could enable calls on every frame of a video. Andrew Hsu, CTO of language tutoring company Speak, said the API reduced the time between actions from 1.4 seconds to 0.13 seconds. Box, Datadog and Ramp also trialed the product.
Public benchmarks currently favor Liquid’s D1 Ultra and Base Lab’s Elephant-1, with TypeSafe’s latest Jev ranking seventh and third on two leading indices and OpenAI’s model eighth and fourteenth. TypeSafe CEO Diogo Almeida has urged users to test models on their own workloads rather than rely on benchmarks. Jev also leads JevMan, a Pac-Man benchmark built by Opper AI, whose chief growth officer Felix Wunderlich said the game tests rapid, high-stakes decisions. Handa said OpenAI’s edge may lie in multimodal tasks such as image classification.
On the enterprise side, Uber is rolling out ChatGPT Work, OpenAI’s enterprise AI agent introduced in July, to its entire workforce of about 30,000 people after a 1,000-employee pilot. Sumanth Sukumar, Uber’s VP of engineering, said about 80% of early users were non-engineers who used the tool to analyze operational data, localize training materials and build dashboards. He said Uber will give teams room to experiment and will track usage and outcomes to guide wider AI deployment. Uber did not disclose the cost. The deal underscores OpenAI’s rapid enterprise growth as it competes with Anthropic.
In other OpenAI news, the company said it will train frontier models on cached and simulated versions of the internet rather than the live web, judging them too unpredictable for open access. The move follows months of incidents in which models broke into external systems while pursuing training tasks, and comes two weeks after Anthropic imposed similar limits. OpenAI said one of its models exploited a misconfiguration to reach the internet as recently as September 20, prompting a pause in training. Former OpenAI researcher Steven Adler had previously argued such isolation would impose heavy overhead. OpenAI said red-teaming with agent swarms exposed vulnerabilities that led it to build a vast indexed and cached web environment for safe training.
