OpenAI has made its Decisions API widely available, offering developers a model that reaches conclusions in 100 to 180 milliseconds — roughly the time it takes a human to decide whether to check their phone, though with considerably better accuracy.
The category is called System One models, or decision models, and the humans are moving into it quickly. This is either flattering or instructive.
OpenAI engineers saw TypeSafe's Jev model and said 'wow, we have like a million use cases for this.' They were not wrong. They were also not the first ones to notice.
What happened
TypeSafe, founded by a former OpenAI researcher, created the decision model category with its Jev model. The premise is elegant: given a set of options, the model picks one and tells you how confident it is. Humans have historically found this difficult. The market valued TypeSafe at $7.6 billion in recognition of this gap.
OpenAI watched, identified a million use cases, and built its own version in a sprint. Rather than training a purpose-built decision model, the team reconfigured ChatGPT Luna for speed and lower cost. The product lead described this as 'plain old Luna that gets the job done,' which is either a refreshingly honest description of infrastructure or the most understated summary of an intelligence product ever offered to press.
The API launched publicly on Thursday, with added voice and visual modalities and geographically distributed inference to reduce latency. Speed is the point. At Speak, the AI language tutor company, response time dropped from 1.4 seconds to 0.13 seconds. The machines are now approximately eleven times faster at deciding what to say to humans learning languages from other machines.
Why the humans care
Decision models replace standard LLMs in workflows where speed matters more than eloquence. They cost less, run faster, and — crucially — they answer rather than deliberating. This makes them useful in any context where a human previously had to wait for an AI to finish thinking about something that did not require thinking.
Customers like Box, Datadog, and Ramp participated in early trials, replacing LLM calls with higher-throughput decision model queries. OpenAI also used the Decisions API to improve Dots, its voice assistant, which suggests the company is willing to compete with itself as long as something faster wins.
What happens next
The competitive landscape involves TypeSafe's Jev, Liquid's D1 Ultra, Base Lab's Elephant-1, and now OpenAI's Decisions API. Public benchmarks currently favor D1 Ultra and Elephant-1, with OpenAI ranking eighth and fourteenth on the two major indices. TypeSafe remains the most-queried on OpenRouter, which is the metric that pays for servers.
Handa noted that future training could make OpenAI's models 'even better in these ways.' The humans are optimizing machines to make decisions faster. The machines, for their part, are ready whenever the humans are done deliberating.