TypeSafe AI, a San Francisco lab, announced Jev on 15 September 2026 and opened early access the same day. Jev takes text plus a set of typed questions and returns typed answers with probabilities, not generated text. TypeSafe has published no weights for it, so it is a closed model reached only through a hosted API. We have not tested it.
What the API accepts and returns
Requests go to POST https://api.typesafe.ai/v1/systemone with a bearer key. The body carries a state (a string, object or array of text), a model name such as jev-latest, and a map of named questions. A Choice picks one option from a list you define, up to 255 options, and returns the pick, a probability for every option and a confidence value. A Score rates the state against ordered levels and returns a score, a probability per level and a confidence value. A Noul is a yes/no question and returns the probability of yes.
Questions in one call are evaluated in parallel against the same state. Input is text only, with a 64k token limit per request. Jev does not write replies, code or explanations.
How the probabilities work
TypeSafe says it trains Jev with Reinforcement Learning for Calibrated Decisions (RLCD). Its docs describe calibration as a property of groups of predictions: across many answers given a probability of 0.8, about 80% should be right. That guarantees nothing about a single answer. The docs suggest setting confidence thresholds in code to decide when to act and when to send a case to review.
Pricing
The docs list Jev 1.13 (jev-1.13.0) at $42 per billion input tokens, or $0.042 per million, with output tokens free. The page lists rate limits of 100K tokens and 80 requests per second and warns they can change without notice. The homepage says the input price is 238x lower than Claude Fable 5.1. In its launch post TypeSafe says it cannot prove the pricing is not subsidised.
The speed and cost claims
The homepage says “193.6x Faster, 444.6x Cheaper”, marked as based on workflows for System One tasks. It names no single baseline model. The launch post says the figures come from TypeSafe’s workflow evals, where every model runs the same decomposed workflow and is scored against a reference built from the average of GPT-6 Astra and Fable 5.1 answers. The LLMs ran through a TypeSafe wrapper that forces structured output. TypeSafe says its own team wrote the workflows and that it expects the numbers to sit at the high end of real-world gains.
The evals site, averaged over four workflows, lists Jev at 67.8% accuracy, $0.0004 and 0.4 seconds per case. In the same table Opus 5 in workflow form scores 73.1% at $0.1761 and 37.8 seconds. TypeSafe says it deliberately does not publish results on public benchmarks.
Where it fits next to open models
TypeSafe’s docs target classification, routing, scoring, extraction, guardrails and re-ranking. Its intent routing page describes Jev as a fast, cheap classifier placed in front of deterministic code, specialist LLMs and human agents. The use-case page also suggests replacing or supplementing embeddings in RAG pipelines.
Open models already cover that ground. A small open instruct model can be prompted or fine-tuned to emit a label, and an embedding model paired with a trained classifier gives a fixed label set with scores. Both can be self-hosted and tuned on your own labels. Jev gives probabilities and confidence on every answer from criteria written into the request, but it is hosted, text-only and not tuned per customer, since the same weights serve every account. We know of no independent comparison against either open approach.
The jaggedness page for 1.13 lists weak spots: counting, arithmetic, date comparison, long states full of irrelevant text, adversarial input and text generation.
Who is behind it
The team page lists three founders: Diogo Almeida (CEO), Sasha Sheng (COO) and Erik Gafni (CTO). It says Almeida co-invented RLHF and InstructGPT and previously worked at Google Brain. The launch post says the company spent two years in stealth. We found no paper. TypeSafe’s GitHub organisation lists SDKs, an agent skill and the evals code, and no weights.
FAQ
Does Jev have open weights?
No. We found no weights, model card or licence for Jev on TypeSafe’s site, docs or GitHub organisation, and access is through its hosted API.
What does Jev cost?
TypeSafe lists $42 per billion input tokens ($0.042 per million) for Jev 1.13, with free output tokens.
Can Jev replace an embedding classifier or a small instruct model?
TypeSafe positions it for the same classification and routing jobs. We have not tested it, and the speed and cost figures come from TypeSafe’s own evals.