What is Jev? TypeSafe's decision model explained
A model that picks instead of writing. What a System One decision model takes in and returns, what it costs, how independent tests rate it, and where it breaks.
LLM Pricing2,602 words10 min read
On this page
Jev is a model from TypeSafe AI that does not write text. You send it a piece of text or JSON, called the state, plus typed questions. For each question it returns a choice, a score or a yes/no probability, together with the probabilities behind the answer. TypeSafe released it in early access on Sep 15, 2026 and calls it the first "System One model". Other labs and the press call the category decision models.
Facts come from TypeSafe's documentation and launch post, two independent benchmark write-ups from Artificial Analysis and Arize, and press coverage, linked under Sources.
Short answer
- Jev is a decision model, not a chat or coding model. It reads text, with images as an opt-in experiment, and returns typed answers with probabilities. It produces no prose, code or explanations.
- A request carries one state and any number of questions. Choice picks one option, Score rates on ordered levels, Noul gives the probability of a yes, and Box (Jev 1.14 only) locates something in an image.
- It costs $0.042 per 1M input tokens and output is free. Paid hosts in this site's catalog list the same rate.
- Jev 1.13 shipped on Sep 15, 2026 and Jev 1.14 on Oct 14, 2026. The weights are closed and the model runs only on hosted APIs.
- Artificial Analysis scored Jev 1.13 at 58 on its Decision Index, first among decision models at launch and 3 points below Claude Haiku 5.5 (Low), at $42 against $184 per million tasks. Newer decision models have scored higher since.
What Jev is
TypeSafe AI is a San Francisco company founded in 2024 by Diogo Almeida, who worked on reinforcement learning from human feedback (RLHF) at OpenAI, with Erik Gafni and Sasha Sheng. TechCrunch reported on Oct 9 that it raised $870 million at a $7.5 billion valuation, a few weeks after launch.
TypeSafe's argument is that chat models are tuned for answers people like to read, which fits badly when software consumes the output. Jev is trained with a method the company calls Reinforcement Learning for Calibrated Decisions (RLCD). The aim is that across many predictions, answers given a probability of 0.8 are right about 80% of the time. TypeSafe has not published the architecture, the weights or a paper. TechCrunch describes the model as transformer-based, and Almeida told it the training data is entirely synthetic.
The names carry two references. "System One" comes from the fast, intuitive thinking in Daniel Kahneman's Thinking, Fast and Slow. "Jev" comes from the economist William Stanley Jevons, whose paradox says that cheaper use of a resource tends to raise total consumption.
How a request works
Everything goes to one endpoint, POST https://api.typesafe.ai/v1/systemone, and the model field picks the version. The body holds a state (a string, a JSON object or an array of text values) and a questions map. Each question has an ID, a type, instructions and, for Choice and Score, criteria that define the options or levels. The ID is for your code. TypeSafe's docs say it is not sent to the model, so the instructions have to carry the whole question.
| Question | Use it for | Answer fields |
|---|---|---|
| Choice | Pick one option from a list, up to 255 options according to TypeSafe's launch post | choice, probabilities, confidence |
| Score | Rate on ordered levels you describe | score (can fall between levels), legend, probabilities, confidence |
| Noul | Decide whether a statement is true | noul, a probability from 0 to 1 |
| Box (Jev 1.14) | Locate one thing in an image | box with four edges as fractions of the image size, confidence |
Every question in a request is evaluated in parallel and independently against the same state, so answers do not feed each other. If a later question depends on an earlier answer, you make a second request. TypeSafe advises one narrow judgment per question and combining the answers in your own code. For example, ask separately about market size, technical feasibility and differentiation instead of "rate this pitch", then weight the scores yourself.
confidence summarizes how concentrated the probabilities are, from 0 for an even split to 1 for all the probability on one option. For a Choice with n options it is (top probability - 1/n) / (1 - 1/n). Noul has no confidence field, because a value near 0.5 already means the model is unsure. TypeSafe suggests three bands: act automatically at high confidence, ask for confirmation in the middle, and route to a person at low confidence, with thresholds that scale with the cost of a wrong action.
Calibration describes groups of predictions. TypeSafe's docs say it does not guarantee that any single answer is correct, and the "cannot hallucinate" claim in its launch post means the answer is one of the options you supplied, not that the answer is right.
What it is used for
TypeSafe's cookbooks cover ticket and intent routing, function calling, guardrails on LLM input and output, citation checking, scoring retrieved passages, re-ranking and hierarchical document classification. Early adopters quoted in the press describe swaps inside existing pipelines. TechCrunch reported a Vercel engineer replacing an OpenAI model in a command-safety classifier and getting results 5 to 18 times faster with higher accuracy. It also quoted the CTO of Bryo AI, who found Gemini slightly more accurate than Jev for email classification but 10 to 20 times more expensive.
Jev does not replace the language model behind a chat assistant or coding agent. TypeSafe's docs say so directly and name Claude Code, Cursor, Copilot, Muse Spark and Grok Bot among the tools it cannot power. The documented pattern is to use an agent to write code that calls Jev for the decisions inside a product.
Versions, images and languages
| Jev 1.13 | Jev 1.14 | |
|---|---|---|
| Model ID | jev-1.13.0 | jev-1.14.0 |
| Released | Sep 15, 2026 | Oct 14, 2026 |
| Question types | Choice, Score, Noul | Choice, Score, Noul, Box |
| Image input | Opt-in header | Opt-in header |
| Default context | 64K per request, 32K per question | 64K per request, 32K per question |
| Long-context mode | 262K | 262K |
| Input price per 1M tokens | $0.042 | $0.042 |
Two aliases point at versions. When this page was read on Oct 23, TypeSafe's Models page listed jev-latest as jev-1.13.0 and jev-preview as jev-1.14.0. Press summaries of the 1.14 release post say jev-latest was due to move to 1.14.0 about a week after Oct 14. An alias moving changes answers without any change in your code, so TypeSafe advises sending a versioned ID if you tuned thresholds against one version. The response's model field records which version answered. Per MarkTechPost's summary of the release post, TypeSafe gives 60 days' notice before retiring 1.13 and 30 days before retiring 1.14. This site tracks the alias and 1.13 routes on one page, Jev, and 1.14 on its own page.
Both versions accept images behind the TypeSafe-Danger-Zone-Images: true header. TypeSafe labels images and Box "dreamer-ready", meaning prototypes only: the request format, answers and limits can change between releases, and accuracy is lower than with text. Images must be PNG or JPEG, up to 16 MiB per file and 280 to 1,568 pixels per side, and they bill as input tokens. A Box question returns four edges every time, even when the target is absent, so TypeSafe suggests pairing it with a Noul that asks whether the target is visible. Audio and video are not accepted.
English is the primary training language. TypeSafe says other languages, including CJK scripts, work but not equally well, and it expected 1.14 to improve most on multilingual input without publishing scores.
Jev price
USD per 1M tokens, read on Oct 23, 2026 from TypeSafe's Models page and this site's catalog.
| Item | Jev 1.13 and Jev 1.14 |
|---|---|
| Input | $0.042 ($42 per billion tokens) |
| Output | Free |
| Images | Billed as input tokens at the same rate |
| Long-context mode | Same per-token rate, slower and less accurate by TypeSafe's account |
TypeSafe's own API is not in the models.dev data this site reads, so the model pages show no official price. Every host in the catalog lists $0.042 input and free output, apart from one route labelled free that lists $0. The catalog's context figures for Jev range from 32K to 262K by host, close to TypeSafe's three published limits of 32K per question, 64K per request and 262K with the long-context header.
Only input is billed, so cost follows the size of the state you send with each request:
| Input tokens per request | Cost per request | Cost per 1M requests |
|---|---|---|
| 500 | $0.000021 | $21 |
| 5,000 | $0.00021 | $210 |
| 32,000 | $0.001344 | $1,344 |
A state is billed once per request however many questions travel with it. TypeSafe's parallel-questions cookbook reports 13 questions in one call as more than 11 times cheaper than 13 separate calls, with the same answers. The launch post also says TypeSafe cannot prove the price is unsubsidized and expects prices to fall rather than rise.
Jev is one of 31 entries of kind decision in the catalog, counting variants and free versions. A few of them, with the lowest listed paid input price and the largest listed context on Oct 23:
| Model | Input per 1M | Output | Context |
|---|---|---|---|
| Jev 1.14 | $0.042 | Free | 262.1K |
| Microsoft Decision-1 | $0.042 | Free | 32.8K |
| Liquid d1-ultra | $0.042 | Free | 32.8K |
| Perplexity Decider V1.2 | $0.02 | Free | 262.1K |
| Inception Mercury Decide | $0.02 | Free | 65.5K |
| OpenAI GPT-6 Luna Decisions | $0.10 | Free | 1.05M |
| Cloudflare Clef | $0.2245 | Free | 65.5K |
OpenAI's Decisions API was announced on Sep 29 and opened to all customers on Oct 22, according to TechCrunch. Arize found that seven of the eight decision models it tested accept Jev's request format. OpenAI's takes one block of plain text and a list of questions instead.
How Jev performs
TypeSafe's launch post claims 193.6 times faster and 444.6 times cheaper than frontier LLMs on its own workflow evaluations, with end-to-end times of 70 to 500 ms. The same post says these gains sit at the high end of real use, that its model-capabilities team wrote the workflows, and that the reference answers are the average of two large LLMs. For 1.14, TypeSafe published charts without a score table, compared only against open-weight models, and said the comparison probably favors its own models.
Artificial Analysis published its Decision Index v1.0 on Oct 13. It averages four tests with equal weight: AA-Omniscience-Decision (knowledge and hallucination), HLE-MultiChoice (reasoning), When2Call (tool calling) and WildGuardTest (safety classification). Results for Jev 1.13:
| Measure | Jev 1.13 |
|---|---|
| Decision Index | 58, ahead of GLiDE and Microsoft-Decision-1 at 54 |
| HLE-MultiChoice | 21, where the other decision models scored 15 or less |
| AA-Omniscience-Decision | Accuracy 35, non-hallucination rate 35 |
| When2Call | 81, fourth place, with the top two at 84 |
| WildGuardTest | 94, one point behind GLiDE |
| Time per task | 0.13 s |
| Cost per 1M tasks | $42 |
Against language models, Claude Haiku 5.5 (Low) scored 61 at 1.4 s and $184 per million tasks, and GPT-6 Luna (Low) scored 55 at 1.5 s and $131. The LLMs led on knowledge: 52 and 66 accuracy on AA-Omniscience-Decision, against 28 to 41 for decision models. Artificial Analysis also found decision models overconfident. Across almost all probability levels, accuracy fell below the stated probability, mostly on knowledge and reasoning questions, while calibration was closer on tool calling and classification.
Arize, an evaluation platform that offers Jev as a built-in judge, tested Jev 1.13 against seven other decision models about three weeks after launch. Jev scored 85.5% balanced accuracy on RAGTruth hallucination detection, 91.1% on 13 Phoenix evaluator suites and 49 of 52 agent routing cases. Its calibration error of 0.023 was the lowest of the eight, though its interval overlapped Liquid d1's. Accuracy among the leaders fell inside each other's confidence intervals. In a wording test, Jev's ROC AUC stayed at 0.93 when the hallucination question was reversed, while several other models fell below 0.5. Arize's advice is to try both phrasings of a question on any model before trusting its scores.
Rankings move quickly. On Oct 22, TechCrunch placed the latest Jev third on the Artificial Analysis Decision Index and seventh on the AIM System One Index, behind newer models such as Liquid d1-ultra, and reported that TypeSafe's models were the most queried decision models on one model gateway.
Limits and data handling
TypeSafe lists these failure modes for 1.13 on a page last reviewed on Oct 2:
- Jev reads questions literally. Write the exact condition and boundary cases in the criteria.
- It does not count reliably and is weak at arithmetic and date comparison. Extract the parts with Jev and compute in code.
- Double negatives and multi-step reasoning cost accuracy.
- Accuracy falls as the state fills with unrelated detail, so filter before sending.
- Text in the state can steer the answer. TypeSafe says Jev does not treat state as hostile by default.
- The order of Choice options can sway the answer toward the first one. Reorder to check.
- It is not trained to generate text.
Rate limits are set per organization, and a request over the limit returns 429. TypeSafe says Jev is not trained on customer requests or responses, and that enterprise customers can arrange zero data retention. The same weights serve every account: Jev is not fine-tuned on customer data, so domain knowledge goes in the state and the criteria.
Frequently asked questions
What is Jev?
A decision model from TypeSafe AI. You send text or JSON plus typed questions, and it returns choices, scores or yes/no probabilities with confidence, instead of generated text. It launched in early access on Sep 15, 2026.
Is Jev an LLM?
No. TypeSafe describes it as a transformer-based model that reads natural language like an LLM but outputs only typed answers and probabilities. It does not write prose or code, hold a conversation or call tools.
How much does Jev cost?
$0.042 per 1M input tokens, or $42 per billion, and output tokens are free. Images bill as input tokens at the same rate. Prices were read on Oct 23, 2026, and the model pages carry the current ones.
What is the difference between Jev 1.13 and Jev 1.14?
Jev 1.14 adds the Box question type for locating objects in images and TypeSafe expects better multilingual accuracy. Both accept images and 262K long-context mode behind opt-in headers, and both cost the same.
What is a decision model?
A model that picks from options you define and reports a probability for each, instead of generating text. TypeSafe calls them System One models. Other makers now ship their own, and this site's catalog lists 31 entries of kind decision.
Can I run Jev myself?
No. The weights and architecture are unpublished, and Jev runs only on hosted APIs. Other decision models, such as Cloudflare's Clef, ship open weights.
Does Jev hallucinate?
Its answer is constrained to the options you supply, so it cannot return a malformed value or an option you did not offer. It can still pick the wrong option, and Artificial Analysis found decision models overconfident on knowledge and reasoning questions.
Sources
Read on Oct 23, 2026. Prices are USD per 1M tokens.
- TypeSafe docs: Models
- TypeSafe: Introducing System One Models and Jev
- TypeSafe docs: System One
- TypeSafe docs: Primitives
- TypeSafe docs: Confidence
- TypeSafe docs: Images
- TypeSafe docs: Jev 1.13 jaggedness
- TypeSafe docs: Jev with coding agents
- TypeSafe docs: Legal
- Artificial Analysis: Decision Index v1.0
- Arize: Decision model benchmark
- TechCrunch: A new kind of AI model from a ChatGPT inventor is thrilling developers
- TechCrunch: The maker of non-text AI model Jev valued at $7.5B
- TechCrunch: OpenAI makes its Jev competitor widely available
- MarkTechPost: TypeSafe AI releases Jev 1.14
- MarkTechPost: TypeSafe AI releases Jev
- OpenRouter: What is Jev
- Wikipedia: Jev (AI model)