OpenAI Decisions API Adds Fast, Structured Choices for Developers

OpenAI is testing a new developer tool designed to make artificial intelligence models faster and more efficient when software needs to choose from a limited set of options.

The company introduced its new Decisions API during OpenAI Dev Day 2026, where CEO Sam Altman described a system that lets developers give the Luna model a predefined set of possible answers. Those choices could include image categories, routing decisions or actions available to an AI agent.

Rather than asking a general-purpose model to generate an open-ended response, the API narrows the task to a specific decision. Altman said that focusing the model on a limited choice can make it significantly faster while retaining capabilities including image understanding, broad language support and safety protections.

OpenAI’s approach resembles TypeSafe AI’s Jev

The concept is similar to Jev, a decision-focused model introduced by TypeSafe AI. Jev is designed for software automation and takes a defined set of possible choices before returning probabilities for those options.

TypeSafe describes the approach as “System One” computing, referring to fast, intuitive decisions rather than the more deliberate reasoning associated with larger AI models. TypeSafe CEO Diogo Almeida, a former OpenAI engineer, joked on X about the start of the “clone wars” after OpenAI disclosed its API.

Almeida also suggested that OpenAI’s interest could reinforce the idea that software will increasingly use specialised, fast decision systems alongside more computationally intensive reasoning models.

The appeal is largely economic and operational. General-purpose large language models can be unnecessarily expensive or slow for software tasks where the possible outcomes are already known. Developers experimenting with Jev have reported using it alongside LLMs to reduce latency and cost.

OpenAI has released the Decisions API only as a limited preview, so it is not yet clear how closely its behaviour, pricing or technical design will resemble Jev. Still, developer discussion following the announcement indicates interest in the category.

Decision models could become a security layer for AI agents

One potential application extends beyond classification and routing. Fast decision models may also help monitor AI agents as they carry out tasks.

OpenAI has been developing additional safeguards for autonomous agents, including using separate models to inspect their behaviour. However, running a powerful frontier model for every action can add substantial computing costs.

Shapor Naghibzadeh, a cybersecurity professional leading the startup QueryStory, explored an alternative approach using Jev. He created a Jev Sentinel demonstration that evaluates an agent’s actions against the task it was originally assigned.

The system can block actions that it judges with high confidence to be inappropriate, send uncertain cases for review and allow actions that appear consistent with the agent’s instructions.

Naghibzadeh built the project for a Jev-focused hackathon. According to figures cited in the demonstration, applying this type of monitoring to an agent workflow cost $2.94 using Jev compared with $372 using a frontier LLM.

That cost difference matters because an inexpensive decision model could potentially apply to every action an AI agent takes, rather than only selected checkpoints. Such a system would not eliminate agent failures, but it could provide another layer of review without requiring a full-scale reasoning model for each decision.

A growing category of specialised AI models

OpenAI is not the only company exploring systems that sit between traditional classifiers and general-purpose language models. Other startups are developing tools that deliver structured decisions with lower latency and lower computing requirements.

The central technical question is not simply whether these models can respond quickly and cheaply, but whether their probability estimates remain well calibrated when used in real software environments. Almeida has argued that TypeSafe’s advantage comes from the synthetic data it uses to train systems to produce statistically useful outputs.

OpenAI’s entry gives the approach greater visibility. If decision-focused models prove reliable enough for classification, routing, and continuous agent monitoring, they could become a practical companion to larger reasoning models rather than a replacement.