# Jev Agent (jev-agent.com)

> An independent developer guide to Jev, TypeSafe AI's System One model. Not affiliated with TypeSafe AI. Every figure here was measured against the live API or traced to a primary source; facts last verified 2026-09-19.

Jev answers typed questions about a state and returns structured decisions with calibrated probabilities. It generates no text at all. Every use-case page carries output measured against the live API on 2026-09-18 — the state that was sent, the answers returned with their probability distributions, latency shown both raw and minus a measured network floor, and the cost the API itself reported.

Licence: you may quote any measured figure or fact on this page, with attribution to jev-agent.com. The code samples are free to copy without attribution.

## Key facts

- Vendor: TypeSafe AI. Launched September 15, 2026 with a $40M seed round led by DCVC.
- Co-founder: Diogo Almeida, former OpenAI researcher and RLHF co-inventor.
- Pricing: $0.042 per 1M input tokens. Output is free — no output tokens are generated.
- Latency: 70–500ms end-to-end.
- Context: 64k tokens. The 64k budget covers the state plus every question in the request. A second, tighter 32k limit applies to the state plus the single longest question.
- Rate limits: 250,000 tokens per second and 1,200 requests per minute; either one over the line returns 429.
- Option ceiling: Choice supports a cardinality up to 255 options; beyond that you need a two-stage funnel.
- Modalities: Text only — no image or audio input at launch.
- Open weights: no. Closed managed API only; no self-hosting or VPC option.
- Endpoint: POST https://api.typesafe.ai/v1/systemone. Current version jev-1.13.0; routes jev-latest, jev-preview, jev-1.13.0.
- SDKs: typesafe-sdk (Python), @typesafe-ai/sdk (JavaScript).
- Primitives: Choice returns choice · probabilities{option} · confidence; Score returns score (0..n-1 weighted mean) · legend{index→label} · probabilities{index} · confidence; Noul returns noul (0-1 float) — and nothing else.
- Gateway gotcha: Jev is a decisions model, so OpenRouter's /api/v1/chat/completions rejects it. The working endpoint is /api/alpha/decisions. It is also absent from catalogue listings because its modality is text->decisions rather than text->text, so searching a gateway for "jev" can return nothing while the model is live.

Everything below in one file, for a single fetch: https://jev-agent.com/llms-full.txt

## Jev

- [System One](https://jev-agent.com/system-one): The model class, and why it exists.
- [Benchmarks](https://jev-agent.com/benchmarks): Vendor claims vs independent tests.
- [Limitations](https://jev-agent.com/limitations): When not to use it.

## Compare

- [Jev vs an LLM](https://jev-agent.com/jev-vs-llm): Same decision, both ways.
- [Jev vs a classifier](https://jev-agent.com/jev-vs-classifier): When training your own still wins.

## Try it

- [Chess demo](https://jev-agent.com/chess): Watch Jev pick legal moves, live.
- [Wikirace demo](https://jev-agent.com/wikirace): One Choice hop at a time.
- [Playground](https://jev-agent.com/playground): Run a call without a key.
- [Use cases](https://jev-agent.com/use-cases): Eight patterns with measured output.
- [In an agent](https://jev-agent.com/agents): Tool selection and confidence gating.
- [Built with Jev](https://jev-agent.com/showcase): Verified projects, grouped by pattern.

## Developers

- [API reference](https://jev-agent.com/api-reference): Primitives, shapes and gotchas.
- [Cost & billing](https://jev-agent.com/pricing): What a decision actually costs.
- [Free API key](https://jev-agent.com/api-access): Our endpoint, eight tuned rulesets.
- [Get API access](https://jev-agent.com/getting-started): Waitlist and gateways.
- [Open source](https://jev-agent.com/open-source): Reimplementations and agents.
- [Sources](https://jev-agent.com/sources): Where the facts come from.

## Worked patterns with measured output

- [Support ticket triage with Jev](https://jev-agent.com/use-cases/support-ticket-triage): Ask one Choice question for the queue and one Score question for severity in the same call — both come back in roughly the time one of them would take, for a fraction of a cent.
- [LLM model routing with Jev](https://jev-agent.com/use-cases/llm-model-routing): Put a Choice question in front of your model cascade so a cheap fast model handles the easy majority and only genuinely hard requests reach a frontier model.
- [Agent tool selection with Jev](https://jev-agent.com/use-cases/agent-tool-selection): Let Jev choose the tool and let an LLM fill in the arguments — the selection step stops scaling badly as your tool count grows.
- [Content moderation with Jev](https://jev-agent.com/use-cases/content-moderation): Use Score for severity tiers and Noul for specific policy checks, then act on the probability rather than a hard label.
- [Phishing and email security classification with Jev](https://jev-agent.com/use-cases/phishing-detection): Jev's single raw verdict lost on accuracy to a cheap LLM in an independent test — but combining five cheap signal questions from the same call beat it decisively.
- [RAG passage classification and reranking with Jev](https://jev-agent.com/use-cases/rag-reranking): Score each retrieved passage for actual relevance to the question and drop the ones that only matched on embedding similarity.
- [Agent output guardrails with Jev](https://jev-agent.com/use-cases/agent-output-guardrails): Ask several yes/no verification questions about a draft response in one parallel call, and block or regenerate when any probability crosses your line.
- [Lead scoring and qualification with Jev](https://jev-agent.com/use-cases/lead-scoring): Score fit and intent against your own rubric, and route the hot ones to a human immediately rather than in tomorrow's batch.

## Start here

- [Get API access](https://jev-agent.com/getting-started): Waitlist, gateways that already carry Jev, and a first request.
- [Official documentation](https://docs.typesafe.ai): The vendor's own docs. Where this site and the official docs disagree, the official docs are correct.

## Optional

- [Sources and editorial policy](https://jev-agent.com/sources): Where every fact here comes from and how it is checked.