The OpenAI Decisions API is now generally available, giving developers a faster and cheaper way to handle routine AI decisions. Announced Oct. 22, the release expands competition with TypeSafe’s Jev and other specialized decision models.
Unlike traditional language models that generate lengthy responses, the Decisions API selects answers from predefined options. It uses GPT-6 Luna, optimized for lower latency and structured outputs rather than extended reasoning.
OpenAI says the system can operate up to 10 times faster than Luna through its standard Responses API. The public release adds image inputs, native JSON support and integration with GPT-Live for real-time voice applications.
OpenAI Decisions API Targets Faster AI Workflows
The service supports classification, request routing and action selection for software agents. Developers can receive probability estimates, predefined choices or numerical scores based on their applications’ requirements.
Early customers include Box, Datadog, Ramp and Speak. Language-learning company Speak reported reducing the delay between actions from 1.4 seconds to 0.13 seconds during testing.
Box reported approximately fivefold faster document-routing decisions without losing accuracy in its limited evaluation. Ramp said the API reduced its AI cost attribution expenses by 30% compared with standard Luna processing.
OpenAI previously introduced the technology during its September developer event. TechAmerica.ai covered the initial Decisions API preview and its potential applications in automation and agent monitoring.
The service costs $0.10 per million input tokens, with discounted rates for cached inputs. Developers can submit up to 400 questions per request.
TypeSafe Jev Expands as Decision Model Competition Grows
TypeSafe helped popularize specialized decision models with Jev, which focuses on rapid classification and probability-based choices. Its latest Jev 1.14 release introduced experimental image support, longer context options, and additional structured-output capabilities.
The company cautions that some newer features are not yet suitable for production use. TypeSafe has also gained developer traction, with Jev adoption on Vercel reaching 35.3% of paid gateway teams within two weeks.
Independent comparisons show that no single model dominates every workload. The AIM System One Index places Liquid’s d1-ultra among leading models, while Jev 1.14 ranks seventh.
The Artificial Analysis Decision Index evaluates classification, reasoning, knowledge and tool-calling performance. Rankings differ because the evaluations measure distinct capabilities rather than one universal measure of intelligence.
Usage also tells a different story. OpenRouter’s decision model rankings show TypeSafe’s Jev models leading request volumes, despite competition from OpenAI, Liquid and other developers.
Speed and Reliability Remain Key Challenges
Developers must balance response time, cost, accuracy and confidence calibration when selecting decision models. A faster answer is useful only when the model makes sufficiently reliable choices for its intended task.
Opper AI’s JevMan benchmark takes an unusual approach by testing decision models in Pac-Man. The evaluation measures how quickly and effectively models choose actions during continuous gameplay.
OpenAI expects further latency improvements to enable more frequent AI decisions, including real-time visual applications. Faster systems could eventually process individual video frames or coordinate numerous agent actions.
The broader economic argument resembles Jevons paradox, where greater efficiency can increase overall resource consumption. Cheaper AI decisions may encourage developers to automate tasks previously considered too expensive.
For businesses, the practical question remains whether specialized decision models deliver dependable results at lower operating costs. OpenAI’s wider release gives developers another option to evaluate against established competitors.
