Most AI headlines are about models that think harder. Jev, from TypeSafe AI, is built to decide faster, and for a lot of real software that’s the more useful trick.
Thinking fast and slow, for software
Psychologists describe two modes of thought. System Two is slow and deliberate: working through a problem step by step. System One is fast and automatic: the instant judgements we make hundreds of times a day.
For the past few years, AI progress has mostly meant better System Two: models that reason longer and write more. But much of what software actually needs is System One. Is this support ticket about billing or delivery? Is this request urgent? Should this agent be allowed to run that command?
On 28 September, TypeSafe AI introduced System One models, a new class of model built for exactly those decisions, and released the first of them, Jev. TypeSafe was founded by Diogo Almeida, previously a researcher at OpenAI, who framed the problem simply: models have been superhuman at chat for years, so where is all the automation?
What Jev actually does
Jev doesn’t return paragraphs. You send it some context, as text, JSON or a list of strings, and a set of typed questions. It returns structured answers with calibrated probabilities that your code can act on straight away. There are three kinds of question:
- Choice: pick from a set of options, such as which team should handle a ticket
- Score: rate something on an ordered scale, such as severity or priority
- Noul: a yes or no judgement with a probability, such as whether a message asks for a refund
Because every answer has to be one of the allowed values, there’s no free text to parse and no made-up categories to catch. Questions are evaluated in parallel, so asking several barely changes the response time.
The numbers
TypeSafe’s launch figures are striking, and worth reading with the usual care for vendor benchmarks:
- End-to-end responses in around 70 to 500 milliseconds, against seconds or minutes for typical LLM calls
- Roughly 40 to 200 times faster than existing LLMs on these tasks
- Input priced at $0.042 per million tokens, with output free
- Reported production gains of 193.6× in speed and 444.6× in cost
TypeSafe itself notes that its workflow results reflect the higher end of real-world gains. Its own guidance is clear that the latency range is product positioning rather than an SLA, and that probabilities are signals, not guarantees of accuracy.
Where it fits
The use cases are wherever software makes the same kind of judgement again and again:
- Triage and routing. Classifying support tickets, emails or documents and sending them to the right place.
- Model routing. Deciding whether a request needs an expensive reasoning model or a cheap, fast one.
- Guardrails for agents. Checking a proposed tool call for risk before it runs. LangChain has already built this into its agent harness, using Jev to block risky actions before they execute.
- Prioritisation. Scoring queues by urgency, value or risk so the important work comes first.
It’s not an isolated idea. At DevDay a day later, OpenAI announced its own Decisions API for answering predefined questions with finite answers. When two very different companies ship the same pattern in the same week, it’s worth paying attention.
Where it doesn’t
Jev isn’t a replacement for an LLM. Open-ended writing, explanation, long-form reasoning and creative work are still System Two jobs. It currently accepts text and JSON, not images, audio or video. And complex decisions need splitting into smaller, well-defined questions that your code combines.
How we’d adopt it
The pattern we expect to see in well-built AI products is simple: System One for the many small decisions, System Two for the few hard ones. Here’s how we’d start:
- Pick one bounded, low-risk decision that happens often and is easy to measure, such as ticket routing.
- Test on your own data. High confidence on a benchmark can still miss in your domain or language.
- Set thresholds. Automate the confident answers and send the uncertain ones to a person.
- Keep hard rules for sensitive actions. A probability should never be the only thing standing between an agent and a payment.
- Log everything: inputs, questions, model versions and outcomes, so you can measure and improve.
The prize is software that feels instant, costs a fraction to run and behaves predictably. For many of the AI products we build, that matters more than another point on a reasoning benchmark.
