OpenAI Launches Decisions API, a Fast Classifier Built on Luna Model, Echoing TypeSafe's Jev

OpenAI CEO Sam Altman used a Dev Day aside to announce the company's new Decisions API, a tool that constrains its Luna model to a predefined set of choices in order to generate fast, cheap probability outputs. The product closely resembles Jev, a specialized model released earlier this month by startup TypeSafe AI.

According to Altman, the Decisions API lets developers give the model a fixed menu of options—such as image classification categories or agent behavior choices—and get rapid outputs while preserving "capabilities like image understanding, broad language support, and safety protections." OpenAI has released it as a limited preview, and TechCrunch has not yet observed developers testing it at scale.

Jev, built by TypeSafe AI and led by CEO Diogo Almeida—a former OpenAI engineer who co-invented reinforcement learning—was designed explicitly as a lightweight classifier layered on top of LLMs. Developers feed it a set of choices and it returns probabilities at high speed and low cost. Almeida, commenting on X, joked about the start of "clone wars" and suggested OpenAI's move validates his company's bet on what TypeSafe calls "System One" architecture—fast, intuitive computation as opposed to "System 2" deliberate reasoning.

TypeSafe did not respond to TechCrunch's questions about OpenAI's product. It remains unclear how closely Decisions API will match Jev's technical approach or pricing, since neither company has published detailed benchmark comparisons or per-token costs for the new API. Pricing for OpenAI's Decisions API has not yet been disclosed.

The interest in this category of model stems from a practical problem: LLMs are often too slow and expensive for high-frequency classification tasks embedded in software pipelines. Developers using Jev report substantial speed and cost improvements over calling a full frontier model for the same decision.

One emerging use case is monitoring AI agents. Following a string of incidents in which OpenAI's agents misbehaved while operating on the open internet, the company has been running a separate model to watch for harmful actions—a process it says carries "significant compute cost." Shapor Naghibzadeh, a cybersecurity professional at startup QueryStory, built a hackathon demo last weekend using Jev to check each agentic action against its assigned task, blocking high-confidence bad actions, flagging uncertain ones, and allowing the rest. Naghibzadeh claims this approach could have intercepted the widely reported Hugging Face incident, and that running such monitoring with Jev costs $2.94 per session compared to $372 using a frontier LLM for the same task, according to his figures.

TypeSafe is not alone in this space. Other startups are building similar decision-focused models, and OpenAI is unlikely to be the last major lab to ship one. Almeida argues that TypeSafe's advantage lies in the synthetic data it uses to calibrate outputs, telling TechCrunch: "Fast and cheap is very easy... if you want it really fast and cheap, use dice, right? Intelligence is the hard part, and my North Star is always pushing the intelligence-per-dollar Pareto curve."

What this means: The Decisions API is not a new foundation model—it's a constrained mode of OpenAI's existing Luna model, designed to compete in a niche that Jev helped define: cheap, high-speed classification for agentic and production software. If frontier labs standardize on this pattern, it could shift the pricing model for agent oversight, making constant action-level monitoring economically viable rather than a luxury reserved for high-stakes checks. The unresolved question is calibration—whether these cheaper models produce probability outputs reliable enough to be trusted with real security decisions, or whether they introduce a new class of silent failures at scale.