Jev by TypeSafe AI

What is Jev?

Jev is TypeSafe AI's System One model for developers who need structured judgments inside software. Instead of writing a response for a person to read, it evaluates supplied context against typed questions. The resulting values can drive routing, filtering and other application decisions.

Three kinds of decision

• Choice: select from defined options, with probabilities and confidence.
• Score: evaluate a specific rubric, returning a score, probabilities and confidence.
• Noul: evaluate a statement with a value from 0 to 1; this result has no separate confidence field.

For example, a support system could ask which queue fits a message, then separately assess whether it needs human attention. Keep each question focused and combine decisions in application code.

Practical application ideas

The following examples describe possible designs, not measured product outcomes.

• Support triage: distinguish billing questions, technical problems and general inquiries. A separate component or human agent writes the response.
• Agent routing: decide whether a request calls for search, document retrieval or review. The application must still validate permissions and arguments before running a tool.
• Directory curation: propose categories for product descriptions and flag uncertain records for editorial review. A category prediction does not establish that a vendor's claims are true.
• Retrieval filtering: score candidate documents for relevance before passing selected material to an answering system.
• Feedback organization: separate feature requests, bug reports and usage questions. Multiple independent questions can help preserve mixed intentions instead of forcing everything into one label.

A sensible integration workflow

1. Define the exact decision your application needs and what should happen when evidence is insufficient.
2. Prepare relevant text as the state. Write explicit options or criteria instead of relying on vague instructions.
3. Call the official API or a supported SDK and read the typed result.
4. Evaluate real examples, including ambiguous and adversarial inputs. Set automation thresholds using those results.
5. Record the model version and review outcomes after changing the model, questions or business rules.

For a tool directory, an editor might separately ask whether a description concerns audio processing and whether it explicitly mentions transcription. Code then combines those judgments into a proposed category. Missing evidence can leave a record in review. This is an illustrative workflow, not a claim that Jev currently performs ToolAI's publishing operations.

Pricing and supported input

Checked September 24, 2026: official documentation lists jev-1.13.0 at $0.042 per million input tokens, with no output-token charge. Free output does not mean free API usage. Input is text only. The documented budgets are 64k tokens per request and 32k for state plus the longest question. Consult the current official specifications before integration.

Limitations and model selection

A typed answer can still be wrong. The documented weaknesses include precise arithmetic, date comparison, indirect reasoning, distracting context and adversarial inputs. Keep exact calculations in code and test borderline examples. English is the primary language; validate Chinese and other languages on representative data.

Use a generation model for article writing, email drafting, code generation or open-ended conversation. Jev can serve as a decision component alongside it. Open-source client libraries do not establish that model weights are available for self-hosting.

Official references

Introduction · Models and pricing · Known limitations. This profile combines verified documentation with illustrative integration advice; it is not an independent benchmark.

Explore related resources

Jev model details · TypeSafe Python SDK · TypeSafe JavaScript SDK · TypeSafe Agent Skills