What Jev is
Jev is a model from TypeSafe AI that answers typed questions about text. It was announced on 15 September 2026, as the company came out of stealth with a $40M seed round led by DCVC.
It is the first model in the category TypeSafe calls System One models: models built to make fast, structured decisions that software can use directly. Jev does not write replies, produce code, or explain its reasoning. Ask it something and you get back a value with a type, not a sentence.
What it returns
You send two things: a state (the content you want judged) and a set of questions. Each question is one of three types.
Choice picks one option from a list you define. You get the winning option, a probability for every option summing to 1, and a confidence value from 0 to 1. Up to 255 options per question.
Score places the content on an ordered scale you describe in words. Minimum 2 levels, maximum 10. The number you get back is the probability-weighted mean of the level numbers, so a distribution of 70% on level 1 and 30% on level 2 returns 1.30, not 1 or 2.
Noul asks a yes/no question and returns one number between 0 and 1: the probability the answer is yes. No separate confidence value comes back, because that number already is the certainty measure.
A ticket classified as “returns” with all the probability mass on that option comes back as confidence 1.0. The primitives guide walks through each shape with real output.
Versions and aliases
The current model is jev-1.13.0. Two aliases point at it: jev-latest and jev-preview. Both resolve to jev-1.13.0 today.
TypeSafe publishes a per-version page of known weaknesses under the name jaggedness. For 1.13 that list covers nine areas, including literal reading of questions, arithmetic and counting, and date comparison. Pinning the exact version rather than an alias means that page stays accurate for what you are actually running.
Pricing
Input tokens cost $0.042 per million ($42 per billion). Output tokens are free.
Free output is not a rounding trick. Jev’s outputs are typed values and probability distributions, so there is almost nothing to bill for. The cost of a call is the cost of the state you send plus the text of your questions. Working out your own bill has a worked example.
Speed, as TypeSafe reports it
TypeSafe’s launch post claims 70ms to 500ms end to end, against “3 to 329 seconds” for frontier models on the same tasks, and describes Jev as “40x-200x faster for the same levels of frontier intelligence.”
Those are vendor numbers from TypeSafe’s own evaluations. No independent benchmark existed as of 18 September 2026, the comparison set is not published in detail, and the frontier-model range is wide enough to cover very different workloads. Treat the direction as credible and the multiple as unverified.
Limits
Total context is 64k tokens per request. Within that, state plus the longest single question can run to 32k tokens. Rate limits are 250,000 tokens per second and 1,200 requests per minute.
Input is text only: a string, a JSON object, or an array of text values. No image, audio or video input.
The maximum number of questions per request is not documented anywhere in TypeSafe’s docs. Since all questions in a call are evaluated in parallel, the practical ceiling is the 64k context budget.
Access and SDKs
Jev is a closed managed API in early access, opened off a waitlist. The weights are not released, so there is no self-hosted option.
The endpoint is POST https://api.typesafe.ai/v1/systemone, authenticated with an Authorization: Bearer header. Both SDKs read the key from TYPESAFE_API_KEY.
pip install typesafe-sdk
npm install @typesafe-ai/sdkThe Python SDK is typesafe-sdk and needs Python 3.10 or newer. The JavaScript SDK is @typesafe-ai/sdk and needs Node.js 20 or newer. Getting access and making your first call covers the console, the key and a working request.
What Jev is not for
Anything that ends in prose. Jev 1.13 “is not trained to generate text,” in TypeSafe’s own words, so drafting, summarising and code generation are out.
Numeric work goes the same way. The jaggedness page lists arithmetic, counting and date comparison as weak areas, and says the model “reads dates as text, not as ordered quantities.” Counting occurrences in a long document, checking whether a date falls in a window, or summing line items should stay in your code.
Adversarial input needs a screen in front of it. “State is data, and jev-1.13 does not treat it as hostile by default.” If user-supplied text reaches the state, a separate prompt injection screen is your job, not the model’s.
Judgments that need a chain of inferences go badly as well. Multi-hop indirection and double negatives both reduce accuracy, so break the judgment into narrow, independent questions, which is what the build guide is about.
FAQ
What does Jev actually do?
Jev takes text you supply and answers typed questions about it. It classifies into options you define, scores against scales you define, and returns probabilities for yes/no questions. Every answer arrives with a probability distribution, and Choice and Score answers also carry a confidence value from 0 to 1.
How much does Jev cost?
Input tokens cost $0.042 per million, which is $42 per billion. Output tokens are free. Your bill is driven entirely by the size of the state you send plus your question text, so trimming irrelevant context cuts cost and, per TypeSafe’s docs, improves accuracy at the same time.
Is Jev open source?
No. Jev is a closed managed API in early access and TypeSafe has not released the weights, the architecture, or the training details. Several community reproductions appeared within days of launch, including openjev and open-jev, but those are built on other models and do not replicate TypeSafe’s method.
Can Jev replace my LLM?
Only for decisions. Jev cannot write, summarise or generate code, so anything producing text for a human still needs an LLM. The common arrangement is Jev making routing and gating calls inside your code, with an LLM handling whatever gets written afterwards.
What model version am I getting?
jev-latest and jev-preview both resolve to jev-1.13.0 today, so all three aliases currently give you the same model. Pinning jev-1.13.0 explicitly is the safer default in production: your confidence thresholds stay tuned against one model, and TypeSafe’s published weakness list for that version stays accurate for what you are actually calling.
How fast is Jev really?
TypeSafe claims 70ms to 500ms end to end and 40x to 200x faster than frontier models. Those figures come from the vendor’s own evaluations and no independent benchmark had been published as of 18 September 2026. Measure it on your own state sizes before designing around a latency budget.