The most talked-about AI launch among developers this month isn't a bigger chatbot. It's Jev, a model from startup TypeSafe AI that deliberately can't write text. Instead, it reads your input and returns a typed decision , a label, a score or a probability , in milliseconds.
Also Read: Build a Multi-Tenant SaaS Panel PHP with Filament v5 (Step by Step)
TypeSafe released Jev 1.13 on September 15, 2026, and on September 20 it dropped the waitlist, so anyone can now sign up and get an API key.
TL;DR
What it is: a "System One" decision model , classification, scoring, yes/no judgments
What it isn't: a text generator. No drafting, summarising or code writing.
Price: $0.042 per million input tokens; output is free
Speed: TypeSafe claims 70–500 ms end-to-end latency
Context: 64k tokens per request (32k for the input "state")
Access: open sign-up at console.typesafe.ai; closed weights, managed API only
Who's behind Jev?
TypeSafe AI was founded in 2024 by Diogo Almeida , a former OpenAI engineer who worked on the training techniques behind ChatGPT , along with Erik Gafni and Sasha Sheng. The company came out of stealth with $40M in funding alongside the Jev launch.
Why "System One"?
The name borrows from psychology's "System 1 vs System 2" idea: fast, intuitive judgement versus slow, deliberate reasoning. LLMs are general-purpose System 2 tools that happen to be used for lots of System 1 jobs , routing a support ticket, flagging spam, picking which tool an agent should call. For those jobs, you pay for a model to write an answer, then you parse and validate the string.
Also Read: News: Claude Opus 5.5
Jev skips the writing. You send text plus a set of typed questions, and it returns structured answers your code can use directly.
How Jev works: Choice, Score and Noul
Every request has two parts: the state (the text to judge) and one or more questions. There are three question types:
Type | What it returns | Example |
|---|---|---|
Choice | One option from a list you define (up to 255), with probabilities | Is this ticket billing , technical or other ? |
Score | A level on an ordered scale (up to 10 levels) | How frustrated is this customer, 1–5? |
Noul | A calibrated probability that a statement is true | Is the customer asking for a refund? |
Multiple questions in one call are evaluated in parallel against the same state. TypeSafe says the model was trained with a technique it calls Reinforcement Learning for Calibrated Decisions, so a 0.8 is meant to be right roughly 80% of the time.
Also Read: News: OpenAI DevDay 2026:
Here's the basic call:
curl -X POST https://api.typesafe.ai/v1/systemone \
-H "Authorization: Bearer $TYPESAFE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "jev-1.13.0",
"state": "I was charged twice this month. Please fix this ASAP.",
"questions": {
"urgent": { "type": "noul", "instructions": "Does this message express urgency?" }
}
}'Speed and cost claims
Tom's Hardware covered TypeSafe's claim that Jev is up to roughly 190x faster and 440x cheaper than frontier models like GPT-6 Astra for this kind of task. Treat those as vendor numbers , but early independent signals are encouraging. Vercel CTO Malte Ubl reported Jev saturating an existing classifier eval that previously ran on Gemini 2.5 Flash Lite, at about six times the speed.
Some engineers are more measured. Sean Goedecke argued that much of the speed may come from constrained single-token output and aggressive prefill rather than an entirely new architecture , while still calling fast structured output an interesting new building block.
What Jev is bad at (by its own admission)
Unusually, TypeSafe publishes a "jaggedness" page listing where Jev 1.13 fails. Highlights:
Literal reading , it answers exactly what you wrote, not what you meant. Avoid double negatives.
Counting and maths , don't ask it to count characters or compare numbers. Do that in code.
Dates , it reads dates as text, not ordered values.
Indirection , tasks that chain two facts together are less reliable.
Prompt injection , it treats input as data and won't defend against hidden instructions. Screen untrusted input first.
Generation , it simply doesn't write text.
Best use cases for developers
Support ticket triage , category, severity and refund intent in one call
Model routing , decide whether a request goes to cheap code, a small model, a frontier LLM or a human
LLM guardrails , check prompts and outputs before they reach users
RAG filtering and reranking , score retrieved passages before they hit your prompt
Agent tool selection , pick the right tool from a long list, fast
Semantic linting in CI , enforce rules like "no business logic in controllers"
The community moved quickly: Browser Use shipped a fast browser agent that uses Jev to choose the action and DOM element, only calling a small LLM when text needs typing.
Also Read: GPT-6.1 Sol: Pricing, Benchmarks Events and Is It Worth Switching?
Should you use Jev?
If your app makes lots of small, repetitive decisions and you're currently paying an LLM to make them, Jev is worth a benchmark this week. Keep an LLM for anything that ends in prose. And pin jev-1.13.0 rather than jev-latest , so the published weakness list matches the model you're calling.
We show a Laravel integration in How to Use Claude, Gemini, Grok & Jev APIs in Laravel .
FAQ
Is Jev an LLM? No. It's a decision model that returns typed answers with probabilities and never generates text.
Also Read: OpenAI Dots Explained: ChatGPT's New Always-On AI Agents - News
How much does Jev cost? $0.042 per million input tokens. Output tokens are free.
Is there a Jev waitlist? Not anymore , TypeSafe removed it on September 20, 2026.
Can I self-host Jev? No. The weights aren't released; it's API-only.
Also Read: DevDay 2026 for Developers: Codex Cloud, Agents API and More
Which languages does Jev support? English works best. Other languages work but TypeSafe advises testing on your own data.
Sources: TypeSafe AI docs (docs.typesafe.ai), OpenRouter, Tom's Hardware, TechCrunch, systemonemodels.org, KDnuggets, Sean Goedecke.
