OpenAI launches Decisions API at DevDay, challenging TypeSafe AI's Jev with faster, lower-cost structured decision-making
October 12, 2026 · 2 min read
Also published in العربية, עברית, Español, Italiano, Português (Brasil)
At its DevDay event, OpenAI introduced the Decisions API, a new interface that allows models to select from a set of predefined answers and return confidence scores. The API is built on GPT-6 Luna, OpenAI's smallest and lowest-cost model, and reportedly delivers responses in 150 milliseconds, compared to 1.6 seconds when using GPT-6 Luna directly for similar tasks.
The Decisions API narrows open-ended generation into a structured choice mechanism, closely resembling Jev's Choice primitive. It enables developers to provide a question, candidate answers, and context, with the model returning a decision rather than generating free-form text. This positions the API as a lightweight component for high-frequency, low-cost integration into agent workflows.
Shortly after the announcement, a Reddit post in the AI_Agents community declared "Jev Is Dead," suggesting that OpenAI's move threatens Jev's market position. Commenters noted that Jev's remaining advantages are speed and cost, and questioned whether it ever had a meaningful competitive moat.
TypeSafe AI launched Jev on September 15, with reports indicating that potential investors were discussing valuations exceeding $10 billion, up from a previous valuation of around $200 million. Jev is positioned as a System One model designed for fast, intuitive decisions, offering three primitives: Choice, Score, and Noul.
TypeSafe claims Jev is 20 to 200 times faster than leading large language models on structured tasks, with costs ranging from 1/40 to 1/400 of comparable models. Input tokens are priced at $0.042 per million, with output tokens free.
However, direct comparisons remain limited to product form and response speed, with pricing, calibration methods, and real-world accuracy still undisclosed. Early clones of Jev have been built on existing models like Qwen3.5-9B and Qwen2.5-0.5B using LoRA fine-tuning and synthetic contrastive data.
Industry observers note that if architecture alone is not a moat, the real barrier may lie in training data and reinforcement learning pipelines. OpenAI's existing developer platform and suite of products like Dots, Space, and Codex could give it an edge in reaching developers once the Decisions API is fully integrated.
TypeSafe founder Diogo Almeida, a former OpenAI researcher involved in GPT-4 and ChatGPT, remains unfazed. He stated in a recent interview that if model quality remains paramount, TypeSafe will maintain a strong position for the foreseeable future.
The Decisions API is currently available in limited preview, with details on pricing, candidate answer limits, and custom data tuning yet to be announced. An OpenAI spokesperson indicated that more information will be shared upon full launch.