What Is Jev? TypeSafe AI's Decision Model Explained
Direct answer: Jev is a probabilistic decision model from TypeSafe AI. It takes state plus typed questions and returns structured decisions with probabilities. It does not write Magicdoor's answers. Magicdoor uses Jev as a routing signal, then DeepSeek V4.1 Flash, GPT-6 Luna, GPT-6.1 Sol, or Claude Opus 5.5 responds.
Fact-checked September 23, 2026. This page owns the Jev and decision-model explanation. See the automatic model selection pillar and exact Magicdoor Auto policy for routing details.
Decision model versus chat model
| Question | Jev decision model | Generative chat model |
|---|---|---|
| Primary job | Evaluate state and return typed decisions | Generate text, code, or other content |
| Output | Choice, Score, or yes/no probability | Free-form token sequence |
| Typical use | Route, classify, score, verify, branch | Answer, explain, draft, reason, create |
| Magicdoor role | Supplies routing probabilities | Writes the user-facing answer |
TypeSafe calls Jev its first System One Model, borrowing the fast, intuitive “System 1” idea from Daniel Kahneman. That is TypeSafe's category and framing, not a claim that every decision is infallible.
What typed decisions mean
TypeSafe exposes Noul for a yes/no probability, Choice for one defined option plus a distribution, and Score for an ordered rubric. The application defines the question and threshold. A Noul value of 0.30 is not “30% difficult”; it is a probability for the stated condition.
Decision models matter for model routing because they return bounded results that software can apply directly. They can also choose a tool, score risk, verify output, or send uncertain cases to review. The surrounding policy still owns thresholds, capabilities, fallbacks, accounting, and observability.
How Magicdoor uses Jev
For each Experimental Auto message, Magicdoor asks whether the latest turn requires substantive inference/judgment or reconsidering a disputed answer, and whether it requests code/data execution.
| Policy result | Answer model |
|---|---|
| OpenAI-required context or execution at least 0.50, with substantive quality or execution need | GPT-6.1 Sol |
| OpenAI-required context without escalation signals | GPT-6 Luna |
| Compatible text with quality probability at least 0.30 | Claude Opus 5.5 |
| Otherwise | DeepSeek V4.1 Flash |
Effective saved file and container history can require an OpenAI route even when the latest prompt is text-only; Jev receives bounded text, not the files themselves. Jev does not choose from a hidden catalog or generate the answer. Magicdoor owns the pool and quality-first-v7 policy. Manual model selection bypasses Jev. See all available models on the models page.
TypeSafe's claims, with attribution
TypeSafe's September 15 launch reports a $0.042 per million input-token price, output tokens described as free, and 70-500 ms end-to-end TypeSafe response time. Its workflow evaluations reached up to 193.6 times faster and 444.6 times cheaper than compared language-model workflows.
Vercel describes Jev as a probabilistic decision model and repeats those figures as claims reported by TypeSafe. TypeSafe also publishes caveats: speed tests were generally run from West Coast laptops, the largest workflow gains are expected to be at the high end, workflow authors can introduce bias, and comparison setup matters.
These are TypeSafe vendor claims, not Magicdoor measurements. Magicdoor does not claim that Auto runs at those speeds, saves those percentages, or produces better answers because it uses Jev.
Magicdoor's evidence is different
Magicdoor ran 164 synthetic scenarios and made 324 live TypeSafe requests. That tested route selection and policy behavior. It did not generate and compare final answers, so it does not show real answer-quality superiority.
Limits and trust boundaries
- Jev is a routing component, not the answer model.
- Typed shape does not guarantee the decision matches user intent.
- Probabilities require policy thresholds and evaluation.
- The 2.5-second deadline can fail closed to Sol for incompatible effective context or Opus 5.5 for compatible text.
- Magicdoor excludes files and tool payloads, but user text can contain sensitive strings or URLs. This is not full redaction.
- TypeSafe created Jev. Its use does not imply partnership or endorsement.
Compare the categories directly: OpenRouter Auto vs Magicdoor Auto, ChatGPT automatic selection vs Magicdoor Auto, the published router ranking methodology, and when a multi-model workspace beats separate subscriptions.
For the application-specific boundary, see what Magicdoor sends to Jev.
FAQ
What is Jev from TypeSafe AI?
Jev is TypeSafe AI's probabilistic decision model. It evaluates supplied state against typed questions and returns choices, scores, or yes/no probabilities instead of writing a free-form chat answer.
Is Jev a replacement for chat models?
No. Jev is designed for structured decisions inside software. A generative model is still needed to write an answer, explanation, or other free-form output.
How does Magicdoor use Jev?
Magicdoor uses Jev to evaluate two routing signals for each Experimental Auto message. Magicdoor's quality-first-v7 policy then chooses DeepSeek V4.1 Flash, GPT-6 Luna, GPT-6.1 Sol, or Claude Opus 5.5 to answer.
Did Magicdoor create Jev or receive a TypeSafe endorsement?
No. TypeSafe AI created Jev. Magicdoor uses its API as one component of Magicdoor's routing policy, which does not imply TypeSafe endorses Magicdoor.
Sources
Accessed September 23, 2026.
- TypeSafe launch for terminology, price, speed, workflow results, and caveats.
- TypeSafe API and Noul documentation.
- Vercel AI Gateway announcement.
- Vercel Jev overview.
Related Resources
Selecting the Right AI Model
Current magicdoor.ai model lineup and practical guidance on when to use each option.
AI Model Costs on magicdoor.ai: Current Chat Rates
Compare current chat-model input, cached-input and output prices, including Claude Haiku 5.5 long-context tiers and Mistral Large 4 guarded sale rates.
What Happened to Perplexity Research on Magicdoor?
The older Perplexity Reasoning and Deep Research picker entries are retired. Learn how to research with current search-enabled chat models.
How Magicdoor Tested Auto Routing—and What the Test Cannot Prove
Read the scope of Magicdoor's 164-scenario synthetic Jev routing study, its development and holdout phases, 324 requests, failures and answer-quality limitations.