Product launch2 min read
Jev: an AI model that doesn't talk – it decides
TypeSafe AI has introduced Jev, a "System One model" that returns typed decisions with probabilities instead of text – in milliseconds and at a fraction of the cost of large language models. For agents and workflows, it is an interesting new building block.
Many steps in agents and automations aren't text tasks at all but small decisions: is this email an invoice? Which tool fits? Is this output risky? Today, a large language model usually handles them – slowly, expensively, and with text you then have to parse. Jev by TypeSafe AI targets exactly that gap.
What Jev is
TypeSafe calls Jev the first "System One model", borrowing Daniel Kahneman's term for fast, intuitive thinking. The idea:
- Typed output. You define a schema upfront – a choice between categories, a boolean or a number. Jev only returns values that fit.
- Probabilities, not certainty. Every decision comes with a calibrated confidence. Your application decides at which level to act or escalate to a person.
- Speed. The vendor reports 70 to 500 milliseconds per answer and claims 40 to 200 times the speed of large language models on such tasks – based on its own measurements.
- Price. According to TypeSafe, input tokens cost $0.042 per million and output is free (as of November 2026).
TypeSafe was founded by Diogo Almeida, who says he worked on the instruction-following methods at OpenAI.
Typical uses
- Routing: sending tickets, emails or leads to the right place
- Agent control: deciding which tool or model comes next
- Guardrails: checking LLM output, detecting jailbreak attempts
- Bulk classification: sorting or scoring large datasets
The division of labor matters: language models keep planning, writing and explaining. Jev takes the many small judgements around them. In a workflow tool like n8n, it would be the classic condition node – but one that understands unstructured input.
What to watch out for
- Early access. Access is currently via a waitlist, and public documentation is still thin.
- Vendor claims. Speed and quality are so far backed mainly by TypeSafe itself. Test with your own data before changing processes.
- Not an all-rounder. No free-form text, at most 255 choice options.
- Price stability. TypeSafe itself says the sustainability of its pricing is unproven.
- Data protection. Inputs go to the vendor's API – check which data you are allowed to process there.
Bottom line
Jev doesn't replace large language models; it complements an underrated part of agent systems: fast, recurring decisions. If the vendor's claims hold up in practice, agents could become noticeably faster and cheaper.