# Jev > The model that started decision models, and changed how I think about putting AI into software. Play with it, don't marry it: early access, closed, with a leak report still unanswered. - Verdict: Watching - Vendor: TypeSafe AI - Layer: [AI](https://cloudlesslabs.com/stack/layers/ai/index.md). Job: The first decision model - Last reviewed: 2 November 2026, by Nicolas Dao (Cloudless Labs, Sydney) - Web page: https://cloudlesslabs.com/stack/typesafe-jev/. The whole stack in one file: https://cloudlesslabs.com/stack/brief.md Jev is why this category exists. It takes your data and typed questions, and returns a probability instead of prose, in about a fifth of a second. That changes how you think about the problem: most AI calls in a business app were never conversations, they were decisions we forced through a chat model. The real value of a hot startup is often that new way of thinking, and Jev delivered that whatever happens to it next. So play with it. Spend a weekend on your own data, and you'll see where decision models work and where they don't. Then look at what you'd be signing up for. It's in early access, the weights are closed, the alias has moved under people, and a leak report has sat for over a week with no account I could find from TypeSafe. That isn't something I'd put in front of a client. What would change my mind: a post-mortem on that report, a published deprecation policy, and a way to run it somewhere other than the US. Until then, I'd use [Baseten's models](https://cloudlesslabs.com/stack/baseten-decision-models/index.md) for real work, and keep an eye on this one. Everything else is in [Which decision model?](https://cloudlesslabs.com/stack/guides/which-decision-model/index.md), including [how it does on images](https://cloudlesslabs.com/stack/guides/which-decision-model/index.md). ## Use it when - An experiment, to see what a decision model can do: there's no quicker way to learn, and the SDKs and docs are good. - Trying images and bounding boxes: Jev 1.13 has the best image score of the hosted models I compared, and 1.14 is the only one with a box. Both are experimental. - Speed on a hosted API matters most: Jev 1.14 is the fastest of them on Artificial Analysis's test. ## Don't use it when - Client production, for now: it's in early access, with no published deprecation policy and no way out if it goes away, because the weights are closed. Use Baseten's models instead. - Personal or confidential data, until TypeSafe explains the leak report of 24 October 2026. Use a model with zero retention by default. - Data that must stay in Australia: it lists US only, and it can't be self-hosted. Use open weights on your own GPU in Sydney. - Text written by people trying to fool you: a planted answer worked 79% of the time on 1.13 and 59% on 1.14 in DecisionRobust. d1-ultra scored 28%. ## Pricing $0.042 per million input tokens, output free, on 1.13 and 1.14. Images are billed as input tokens, with no published rule for how many. ## Gotchas, with fixes - **Gotcha:** A third-party report on 24 October 2026 says Jev's outputs were exposed for every customer. I found no root cause or formal statement from TypeSafe. **Fix:** Keep personal data away from it, and ask TypeSafe for an incident report before any client use. - **Gotcha:** jev-latest moved to 1.14 on 24 October, and 1.14 scores 60.2 on the image benchmark against 72.4 for 1.13. **Fix:** Pin a version, log response.model, and re-run your test set before moving. - **Gotcha:** It's the most overconfident of the decision models on the Decision Index: a stated 0.98 was right about 93% of the time. **Fix:** Set thresholds on your own labelled data, and never read 0.95 as 95%. - **Gotcha:** Images and boxes need the TypeSafe-Danger-Zone-Images header, and TypeSafe says they aren't production-ready. A box always comes back, even when the thing isn't there. **Fix:** Ask a yes-or-no question first, and keep a person behind any image decision. - **Gotcha:** Rate limits aren't published, and they've already been cut once. **Fix:** Read yours in the console, and ask before you depend on them. ## Provisioning No provider ยท No Terraform provider No provider for either. It's an API key and an HTTP call. ## Instead, consider - [Baseten decision models](https://cloudlesslabs.com/stack/baseten-decision-models/index.md): Client systems: a better score, zero retention by default and open weights to leave with. - [Cloudflare Workers AI](https://cloudlesslabs.com/stack/cloudflare-workers-ai/index.md): Your code is in a Worker: Clef Flash answers the same questions with open weights, at a lower score. ## Decision guides - [Which decision model?](https://cloudlesslabs.com/stack/guides/which-decision-model/index.md): Jev, OpenAI Decisions, Clef, Baseten, Perplexity or Liquid? Which decision model fits my job, my volume and my risk? ## The concept behind it - [Ask an LLM, or call a decision API?](https://cloudlesslabs.com/concepts/llm-or-decision-api/index.md): When my app needs a yes, a no, a label or a score, should it prompt a general model or call a model built to decide? ## In The Stack Weekly - [Week 46: Sarvam Prism, and Perplexity's decisions that search the web](https://cloudlesslabs.com/weekly/2026-w46/index.md), watching ## Sources - [Jev models and aliases](https://docs.typesafe.ai/models) - [Jev images](https://docs.typesafe.ai/go/images) - [What Jev is bad at (TypeSafe)](https://docs.typesafe.ai/model-jaggedness/jev-1.13.md) - [TypeSafe: Jev 1.14](https://typesafe.ai/blog/jev-1-14) - [DecisionRobust results](https://yifan-lan.github.io/DecisionRobust/data/asr.csv) ## Related - [AI](https://cloudlesslabs.com/stack/layers/ai/index.md): every tool in this layer, with its verdict - [The whole stack in one file](https://cloudlesslabs.com/stack/brief.md): every use case and tool, for questions that span many - [Map of every Markdown file on cloudlesslabs.com, with what each one answers](https://ask.cloudlesslabs.com)