What is Jev? Jev vs a trained classifier, and when to use each

Explore open-source Laya for a self-hosted decision model with a fine-tuning notebook.

TypeSafe released Jev on 15 September 2026 as its first "System One Model". It calls it "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out."

Instead of writing text, Jev returns structured answers:

• a choice from the options you give;
• a score on a scale;
• a yes/no, as a probability.

It can answer several questions about the same text in one request. For example, for one support email: its category, how urgent it is, and whether it's a complaint.

What TypeSafe highlights:

• Speed. TypeSafe reports responses in about 70–500 ms; actual latency depends on workload and provider. TypeSafe reports large speed-ups over LLMs on these kinds of tasks.
• Price. $0.042 per million input tokens, and output tokens are free (TypeSafe's pricing page, checked October 20, 2026).
• Calibrated probabilities. It's trained with "Reinforcement Learning for Calibrated Decisions", aiming to align confidence with observed accuracy. Check calibration on your own workload.
• Answers that always fit your format. Because the output follows the shape you defined, there's no free text to parse.
• Long inputs. Default limits are 64K tokens per request and 32K per question.
• Languages. English is where it's most accurate. Other languages work, but TypeSafe says "not equally well".
• How you shape its answers. Through the text you send, plus instructions and criteria. TypeSafe doesn't fine-tune Jev on customer data, and says Jev isn't trained on customer requests.

How to get it: use TypeSafe’s API or a provider such as OpenRouter. Mastra documents an integration. Check the model version and availability with your provider.

TypeSafe’s current docs list Jev 1.14 as the newest release, while jev-latest still points to 1.13. Pin a version when testing.

Jev can fit routing, guardrails, evaluation and decisions such as urgency or complaint detection. Your application handles the next action and any review.

Sources: TypeSafe’s launch report and current model documentation, checked October 20, 2026.

A trained classifier starts from your history: a spreadsheet of past items and the answer your team gave each one. It learns patterns and applies them to new items; accuracy on new data can differ from the test results.

That matters when the categories are yours:

• your GL accounts and cost centers;
• your ServiceNow assignment groups;
• your product taxonomy;
• who in your company gets access to which SharePoint site.

A general model needs context to know that "Facilities – East" is where your team sends broken badge readers. Your past tickets do.

Because the model learns from your examples, you can also test it on examples it didn't learn from, and see how often it's right overall and for each category. On the public BANKING77 dataset (77 banking intents), StayCharted models trained on the examples answered 92.6% (AI Classifier Model) and 93.7% (Dedicated AI Model) correctly on held-out rows. See the BANKING77 benchmark and method. These results do not include Jev and are not evidence of an accuracy advantage over it.

Compare the assistant and trained-model benchmark method.

Choose Jev when:

• the categories are generic, and any reasonable reader would agree on the answer;
• you have no labeled examples, and need answers today;
• you're a developer adding a decision to code, an agent or a workflow;
• speed and cost per call matter more than matching one team's habits.

Train a classifier when:

• the categories are specific to your organization, and there's a history of decisions;
• you need to show how accurate it is, category by category, before you rely on it;
• non-technical people need to build, check and improve it;
• you want uncertain answers to go to a person, and their corrections to teach the next version.

Both can be evaluated on held-out examples. StayCharted includes that report in the training workflow; with Jev, you arrange the evaluation. There is no measured Jev-versus-StayCharted winner here.

They work well in sequence:

1. Start with Jev. Ask it to label a few thousand unlabeled rows with your categories. That's fast, and very cheap.
2. Have people check the labels. Fix the wrong ones. This is where your team's way of deciding comes in.
3. Train on the confirmed rows. Now you have a model that answers like your team, with an accuracy you've measured.
4. Keep improving it. On Essentials and Business, uncertain answers go to the Review Queue, and confirmed answers become training data for the next version.

On Essentials and Business, StayCharted's Start from unlabeled rows provides its own grouping-and-review workflow: it groups the rows and suggests labels for you to review before training. Jev is not built into that workflow. Export Jev-labeled rows, verify them, then upload the reviewed file for training.

Checked October 20, 2026. Prices, models and availability can change.

• TypeSafe launch
• TypeSafe models, pricing and data handling
• OpenRouter Jev 1.13
• Jev vs Laya vs a trained classifier
• Claude API vs OpenAI API
• Explore StayCharted