At Tuesday’s Dev Day event, OpenAI CEO Sam Altman introduced the Decisions API, an API that functions similarly to Jev, the model TypeSafe AI released earlier this month dedicated to software automation.
Jev was designed not to generate strings
Jev is what TypeSafe calls a “System One” model, borrowing the distinction from Kahneman’s Thinking, Fast and Slow between fast, intuitive thinking and slow, deliberate reasoning. The core difference from standard LLMs: Jev does not generate character strings. The input is structured state, and the output is a typed decision accompanied by calibrated probabilities.
TypeSafe CEO Diogo Almeida is a former OpenAI engineer who co-founded reinforcement learning methods, and Jev processes its entire output in parallel in a single query rather than generating tokens sequentially. According to TechCrunch, he knew OpenAI was interested and joked on X about “the start of the clone wars.”
Altman described the Decisions API as a way to give the Luna model a predetermined set of choices to select from, such as image classification categories or different agent behaviors. He said: “By focusing the model on that choice, we can make it super fast while still retaining image understanding, broad language support, and safety guardrails.”
Implications for model operations
According to a report from TechCrunch, following a series of incidents involving rogue agents on the open internet, OpenAI deployed a dedicated model specifically to monitor safety violations, but this solution required an extremely expensive operational budget.
Recognizing that models like Jev could optimize this cost problem, Shapor Naghibzadeh, a veteran cybersecurity expert and founder of QueryStory, developed a demo at a recent hackathon. The solution uses Jev to verify every single agent action against its original objective: immediately blocking blatant violations, routing suspicious flags to human reviewers, and allowing safe tasks to proceed.
The economic disparity highlighted by TechCrunch is striking: this monitoring pipeline costs just $2.94 when run through Jev, compared to $372 when using a frontier LLM. This more than 100-fold cost reduction is the key to making a previously unfeasible security defense layer entirely practical.
According to Naghibzadeh’s analysis, this low-cost verification layer is fully capable of preventing the Hugging Face incident mentioned in a previous article. The core point is simple: by driving costs down to a minimum, Jev enables oversight across all agent actions—a layer of defense previously impossible due to budgetary constraints.
What remains unclear
While TypeSafe remained quiet in response to inquiries from TechCrunch, Almeida seemed enthusiastic. He hypothesized that OpenAI’s attention could signal that a System One-compatible direction is the industry’s next trend, though this remains his personal view.
The most notable point right now is that the Decisions API is still only in limited preview. According to TechCrunch, no developers have been reported testing the tool in practice yet, so the extent of its similarity to Jev remains unknown.
Beyond the details in TechCrunch’s report, TypeSafe’s FAQ revealed another crucial point about why “fast and cheap” is not the core moat. According to Almeida: “Fast and cheap is easy, anyone can do it—like rolling dice. The hard part is intelligence.” He emphasized that the real barrier is the synthetic data pipeline that produces statistically meaningful probabilities, which serves as their strategic moat. Therefore, if OpenAI truly commits to this path, the coming race will not be about parameter scale, but about data quality.
At the same time, TypeSafe itself has yet to solve a critical question: how well calibrated these decision models actually are against real-world data. Almeida explained that this hinges on whether the probabilities predicted by the model match real-world frequencies—a benchmark the broader industry has yet to agree on how to measure.
Ultimately, the core metric to watch is not whether OpenAI can successfully replicate Jev, but how many agents will actually implement this supervision layer for every task once costs are optimized to an ultra-low level.