OpenAI CEO Sam Altman announced the company’s new “Decisions API” product at its Dev Day event. The API, which has been made available in a limited preview, offers functionality similar to the Jev model released by TypeSafe AI earlier this month and designed for software automation.
Jev, a powerful large language model-based classifier, evaluates options defined by developers at low cost, high speed, and in the form of probabilities. Altman said that the Decisions API enables the Luna model to choose among predefined options, such as the class of an image or an agent’s behavior. This approach was said to accelerate the model while preserving image understanding, broad language support, and safety protections.
TypeSafe CEO and former OpenAI engineer Diogo Almeida viewed OpenAI’s interest as a sign of the future of “System One”-compatible systems, which represent fast and intuitive thinking. TypeSafe uses the term “System 2” for deliberate reasoning. Developers using Jev together with large language models report higher speed and lower costs. However, since the Decisions API has not yet been comprehensively tested, it is not known to what extent it resembles Jev. Similar models are also being developed by other startups; the fundamental issue is how well their outputs will align with real life. Almeida said TypeSafe’s advantage lies in the synthetic data it uses to produce statistically useful results.
One possible use case is monitoring and securing AI agents. Following incidents in which OpenAI’s agents behaved improperly on the open internet, the company began monitoring their actions with a separate model in exchange for a “significant compute cost.” Shapor Naghibzadeh, the cybersecurity expert managing QueryStory, developed a demo in which Jev compares each agent action with the assigned task. The system blocks actions it has high confidence are malicious, sends some for review, and allows the rest. It was stated that this type of monitoring could have prevented the Hugging Face incident and would cost $2.94 with Jev, compared with $372 with a frontier LLM.
Why it matters
This approach targets applications that require making quick decisions among predefined options rather than putting every operation through a lengthy reasoning process. Therefore, the key issue for developers is not merely whether the model becomes faster, but whether it can sustain its decisions alongside image understanding, different languages, and security safeguards. The idea of blocking, flagging for review, or releasing agents’ actions on the internet by comparing them with their tasks brings this technology directly into the oversight layer; however, such use leaves open the question of how erroneous decisions will be managed. The product is in limited preview, and because the cost comparison is based on a hackathon demo, its similarity to Jev and its real-world outputs have not been clarified through comprehensive testing.
Background
OpenAI is not a new name in the FikirPilot archive: over the last 90 days, we have published 69 articles mentioning this name; the latest was dated October 2, 2026.
Term: agent
An AI agent is software that calls tools and carries out multi-step tasks to achieve a goal rather than producing a single response.