IA · 23 October 2026 · 5 min read
The Race for «System 1» AI: OpenAI Opens Decisions API to Challenge Jev
In brief: OpenAI has officially launched its Decisions API to all developers, offering a tool engineered for instantaneous algorithmic choices that bypasses the latency of standard language models. The move directly challenges Jev, the decision model developed by $7.6B startup TypeSafe AI. By adapting its Luna model and reducing response times below 180 milliseconds, the lab accelerates an industry-wide transition toward low-latency, non-textual machine intelligence.
by Team Mocchi's
Beyond Language Models: The Rise of Instant Decision-Making
For years, software engineers building AI-driven workflows faced an inherent bottleneck: forcing large language models, designed primarily to generate conversational prose, to perform simple binary evaluations or pick between structured parameters. Prompting a multi-billion-parameter text model to route an incoming support request, triage network traffic, or authorize an internal transaction regularly incurred multiple seconds of latency and substantial token expenses.
This operational friction fostered a new architectural layer known as «System 1» or decision models. Rather than producing natural language token by token, these deep learning networks function as high-performance classifiers: they ingest contextual data, evaluate predetermined choices, and return an actionable selection accompanied by a statistical confidence score. Following the market momentum of TypeSafe AI—the startup founded by a former lab researcher that reached a $7.6 billion valuation behind its textless Jev model—OpenAI has now made its own Decisions API widely accessible to the developer community.
From Seconds to Milliseconds: The Engineering Behind the Decisions API
Originally previewed during the company's DevDay in late September and previously limited to closed beta partners, the public release introduces visual and voice input capabilities alongside geographically distributed inference infrastructure designed to keep network overhead to a minimum.
As reported by TechCrunch, the central engineering breakthrough lies in latency reduction. Nikunj Handa, product lead for OpenAI's APIs, explained that response times have plummeted from typical 1-to-3-second windows down to just 100 to 180 milliseconds. To deliver this speed without waiting for an entirely new foundation model to finish training, the engineering team adapted its Luna model, tuning inference paths to minimize processing overhead and operational cost.
During initial rollouts with organizations such as Box, Datadog, Ramp, and Speak, development teams began swapping out verbose LLMs from latency-sensitive pipeline stages. In workflows like automated moderation and conditional data routing, eliminating text generation removes formatting ambiguities, allowing downstream microservices to ingest decisions almost instantaneously.
The Clash with TypeSafe and the Shift in Software Architecture
The public release of the Decisions API establishes an open battleground between OpenAI and TypeSafe AI. While TypeSafe developed Jev from scratch as a dedicated non-verbal engine focused strictly on machine-to-machine coordination, OpenAI chose a rapid-adaptation approach with Luna while leaving the door open for specialized decision architectures down the road.
This shift carries implications far beyond corporate rivalry: it redefines how distributed software systems incorporate machine intelligence. When an AI component responds within roughly one hundred milliseconds, it ceases to behave as a slow asynchronous dependency and can instead run inside real-time event loops, responsive user interfaces, and multi-agent systems where numerous small decisions must occur in rapid succession without compound delays.
Mocchi's take
For engineering teams and digital organizations integrating artificial intelligence, the general availability of the Decisions API confirms an essential reality: most production software workflows do not require conversational text, but rather fast, economical, and deterministic choices. Restricting full-scale language models to genuine content-generation tasks while offloading routing, validation, and classification to sub-200-millisecond decision layers dramatically cuts compute expenditure and user latency. For companies building custom applications, adopting hybrid architectures that separate generative reflection from instantaneous execution is becoming the most practical baseline for durable, cost-effective automation.