OpenAI Opens Decisions API to All as TypeSafe Jev Rivals Multiply
OpenAI opens Decisions API to all, using Luna for 100 to 180 millisecond choices as TypeSafe, Liquid and Base Lab intensify decision model market rivalry.
Summary
OpenAI made Decisions API available to all customers on Thursday, October 22, 2026, following its September 29 developer day announcement. Instead of training a dedicated decision model, product lead Nikunj Handa’s team reconfigured ChatGPT Luna, reducing inference from one to three seconds to 100, 120 or 180 milliseconds. The release adds voice and visual inputs plus geographically targeted inference. Trials at Box, Datadog, Ramp and Speak replaced OpenAI LLMs for higher throughput; Speak CTO Andrew Hsu cut time between actions from 1.4 seconds to 0.13 seconds, while OpenAI improved its Dots voice assistant. Handa says Luna offers strong image classification and calibration, customers want image outputs, and 60 millisecond latency could support calls on every video frame.
TypeSafe, founded by a former frontier-lab researcher and valued at $7.6 billion, established this System One category with Jev, which rapidly chooses among options and supplies confidence ratings at lower cost than LLMs. Its latest model, released last week with images and smart-if statements, ranks seventh on the AIM System One Index and third on the Artificial Analysis Decision Index; OpenAI ranks eighth and fourteenth. Liquid’s new D1 Ultra and Base Lab’s Elephant-1 lead public benchmarks, while TypeSafe remains OpenRouter’s most queried provider. Jev also tops Opper AI’s PacMan-based JevMan test, built around rapid, high-stakes choices, according to chief growth officer Felix Wunderlich. TypeSafe CEO Diogo Almeida recommends workload trials because intelligence, cost, calibration and speed require different trade-offs.
Positives
- Decisions API reduces Luna inference from one to three seconds to 100, 120 or 180 milliseconds.
- Speak cut the delay between actions from 1.4 seconds to 0.13 seconds.
- Voice, visual support and geographically targeted inference broaden capabilities while lowering latency worldwide.
- Box, Datadog, Ramp and Speak achieved higher throughput by replacing OpenAI LLMs in tested workflows.
- A 60 millisecond target could eventually enable one Decisions API call for every video frame.
Risks & concerns
- OpenAI ranks eighth on AIM’s index and fourteenth on Artificial Analysis, trailing several competing decision models.
- Decisions API currently uses reconfigured Luna rather than a model specifically trained for decision workloads.
- Intelligence, cost, calibration and speed trade-offs vary by workload, leaving no accepted measure of the best decision model.
- TypeSafe remains the most queried decision-model provider on OpenRouter despite OpenAI’s broader release.
- Customers still want more flexibility to use images as Decisions API outputs.