innFactory AI Consulting from Rosenheim is tracking Jev because the model addresses a gap we see in automation projects every day: most decisions in a workflow do not need text, they need a value that software can act on. For a detailed explanation of how Jev differs from an LLM, read our blog post Jev by TypeSafe: the AI model that writes no text. This page summarises the facts from TypeSafe AI’s announcement and documentation. As of 30 September 2026.
What is Jev?
Jev is the first System One model from TypeSafe AI, founded by Diogo Almeida, who by his own account worked at OpenAI on the methods that taught language models to follow instructions. The name “System One” refers to Daniel Kahneman’s distinction between fast, intuitive judgement (System 1) and slow, deliberate thinking (System 2): reasoning models are System 2, Jev is meant to be System 1.
The vendor’s short formula: a function call with frontier intelligence. Unstructured state in, typed probabilistic decisions out. Jev generates not a single token of text, so sequential generation disappears and all questions in a call are answered in parallel.
How a call works
A request to POST /v1/systemone contains a state (text, JSON object or text array), the model (jev-latest) and a map of questions. There are exactly three question types:
| Type | Answers | Returns | Limit |
|---|---|---|---|
| Choice | Which of these options applies? | chosen option, probability per option, confidence | up to 255 options |
| Score | Which level on an ordered scale? | value, probability per level, confidence | 2 to 10 levels |
| Noul | Is this statement true? | probability of “yes” between 0 and 1 |
The response carries the same keys as the questions, plus a model field with the version actually used (currently jev-1.13.0) and usage with input and output tokens. Customisation happens through state, instructions and criteria in the request, not through fine-tuning. Error codes: 401 (key), 422 (validation), 429 (rate limit), 529 (overload).
Pricing, limits and availability
- Price: USD 0.042 per 1 million input tokens; output tokens are not charged (per the docs, “too cheap to meter”).
- Context: 64,000 tokens combined per request, of which at most 32,000 for the state plus the longest question. Text input only.
- Rate limits: 100,000 tokens per second and 40 requests per second, dynamic and changeable without notice per the docs; higher limits on enterprise plans.
- Latency: 70 to 500 ms end to end per the vendor; TypeSafe claims 40 to 200 times faster processing than frontier LLMs on comparable System One tasks and, in a workflow example, a factor of 193.6 (time) and 444.6 (cost). These are vendor figures without independent measurement.
- Access: early access with waitlist since 15 September 2026; SDKs for Python and JavaScript, an agent skill for Claude Code, Codex and other agent environments.
- Languages: English as the primary language; other languages including CJK with variable accuracy per the docs.
Data protection and operation
Jev is available exclusively as a hosted API. No EU region is documented; per TypeSafe the service is operated from the US West Coast. For the GDPR assessment that means a third-country transfer. TypeSafe provides a data processing agreement with EU standard contractual clauses, states in its docs that Jev is not trained on customer requests or responses, and offers enterprise customers zero data retention. The public docs do not mention security certifications such as SOC 2. Before production use with personal data, the DPA, retention periods and purpose of processing belong in the data protection impact assessment.
Positioning: classifier, not LLM
Jev cannot invent a value outside the schema; in that sense there is no hallucination and no type error. The model can still be wrong: it then picks the wrong option. TypeSafe itself writes that calibration is measured across many predictions and is no guarantee for a single answer. The confidence it returns therefore belongs in every decision logic: automate high confidence, escalate low confidence to people.
Jev suits fast, narrowly scoped judgements inside software: classify, route, score, check, set guardrails, turn large data volumes into features. It does not suit anything that needs text as a result. That remains the job of LLMs such as GPT, Claude or Gemini; in many architectures a System One model sits as a router in front of the expensive reasoning model.
Open alternatives: Laya and Clef
If you want to use the System One idea but keep data in the EU, there are two open options with a Jev-compatible endpoint. Laya by Convai Innovations is a small encoder model (322 to 421 million parameters) under Apache 2.0 that usually needs fine-tuning for good results. Clef and Clef-flash by Cloudflare (since 1 October 2026, 27 and 9 billion parameters on Qwen bases, also Apache 2.0) additionally process images and video, run hosted on Workers AI or self-hosted, and per Cloudflare lead Jev on 7 of 10 decision benchmarks; Jev leads on knowledge benchmarks. Details and each vendor’s published comparison figures are on the model pages.
Our recommendation
Jev pays off where an LLM is currently misused as a classifier, judge or router and latency or cost hurt. Check three points before adopting it: whether the inputs are free of personal data or the contractual situation (DPA, ZDR) is sufficient for US processing, whether your language is covered well enough, and whether the answer space can really be fixed in advance as choice, score or noul. If you want to know which decisions in your workflows can be offloaded this way, get in touch.