TypeSafe AI: the company behind Jev AI

What is TypeSafe AI?

TypeSafe AI is the company behind Jev AI. Jev AI, also called System One, is their hosted decision model: it returns typed answers with probabilities instead of chat text. This page is an unofficial field guide to how the company, the model, the API, and the public GitHub ecosystem fit together. It is not a TypeSafe product page.

Who is TypeSafe AI?

TypeSafe AI publishes documentation at docs.typesafe.ai and lists models on OpenRouter. Their product thesis: stop asking chat models to decide in prose, and instead call a decision API that returns structured answers with probabilities.

Founder

Public materials name founder Diogo Almeida (OpenAI alumnus). This page summarizes publicly stated company context. It is not an investor memo and does not claim fundraising figures beyond what you verify on primary sources.

Why “System One”?

The name nods to Kahneman’s Thinking, Fast and Slow: System 1 as fast, associative judgment. TypeSafe positions Jev AI as a fast path for classification, routing, and scoring, not as a replacement for long-form generation.

Published comparison numbers

The company homepage highlights 193.6x Faster and 444.6x Cheaper versus prior TypeSafe baselines (TypeSafeAI 1.1 comparison on typesafe.ai). Treat these as vendor claims. Re-benchmark on your traffic.

Official links

• typesafe.ai
• docs.typesafe.ai
• Docs index for agents: docs.typesafe.ai/llms.txt (note: typesafe.ai/llms.txt is not the docs index)
• console.typesafe.ai

Timeline

Public launch window: mid-September 2026. OpenRouter listing for jev-1.13 dated 2026-09-18. Current marketed version string: Jev 1.13.

How the names fit together

TypeSafe AI is the company. System One is the series name for models that return a typed decision instead of a paragraph. Jev AI is the public model in that series. On the direct API, the moving alias recorded here is jev-latest. A response can name a specific build, such as jev-1.13.0. Marketing copy also says “Jev 1.13”. Those strings are not chat-model ids. If you send them to a chat-completions client, the call is the wrong protocol.

The product is a Decisions API. You send state and a questions map. Answers come back under the same keys. There is no assistant message to parse. The three question types are noul (yes or no), choice (one label out of up to 255), and score (an ordered scale of 2 to 10 levels). Choice and score also return a confidence summary. Probabilities are the distribution. What is Jev AI? defines noul in one paragraph. Primitives is the field reference, including the noul data type and a request builder.

Input recorded in this handbook is text only (string / JSON object / array). The context window recorded here is 32,000 tokens. Images, audio, and video are not supported. If a vendor page later says otherwise, trust that page over this one.

Where developers call Jev

Three public paths show up in this handbook. They are not the same HTTP API.

1. Direct TypeSafe. POST https://api.typesafe.ai/v1/systemone with Authorization: Bearer <API_KEY> and body fields state, model, questions. The model string this handbook uses on that path is jev-latest. Keys come from https://console.typesafe.ai. The API quickstart has a Python sketch and the recorded error codes (401 authentication, 422 validation, 429 rate limit, 529 overload). Chat-completions SDKs do not speak this protocol.
2. OpenRouter. The September 2026 channel map records typesafe/jev-latest and typesafe/jev-1.13, with a listing date of 2026-09-18 for jev-1.13. Both are the Decisions API, not chat completions. Jev AI on OpenRouter repeats only those ids and tells you to re-check the live card. This page does not add a third id or an OpenRouter HTTP path.
3. Vercel AI Gateway. Dated changelog entries live on the September 2026 channel map. Use that note for Gateway model ids. Do not copy a price off a blog sentence without opening the changelog again.

What a call costs

Public list price recorded here is $0.042 per million input tokens, which is $42 per billion. Output tokens are free. A worked example in the fact card is about 300 input tokens, or $0.0000126 per request. Claimed P50 latency is about 0.23 seconds. The company homepage also publishes 193.6x Faster and 444.6x Cheaper against a prior TypeSafe baseline (TypeSafeAI 1.1 comparison on typesafe.ai). Those are vendor claims. Run your own volume through the pricing calculator before you commit a budget. The calculator does not call the API and does not ask for a key.

The official cookbook says 13 questions in one request can be about 12.2× cheaper and 10× faster than asking them separately, with the same answers. Batching is the normal posture. Keys inside the questions map are your names. They come back on the answers. They are not part of the reasoning.

GitHub projects, skills, and cases

The GitHub projects page is a directory of public repositories that call Jev or document it. It is not an official TypeSafe catalog, even when it links an official SDK. Counts on that page are computed from the data files in this repo, so they move when the daily intake adds a repository. “Jev skills” in search usually means agent skills, not model weights. Jev skills on GitHub groups every catalog entry whose name, repository, or summary mentions skill.

Featured README note for typesafe-ai/skills: The README describes a typesafe-ai skill you can install as a Claude Code plugin or with npx skills add. It is for designing TypeSafe workflows, finding current docs and cookbooks, and composing typed judgments in code. The listed license is MIT.

Community posts are not documentation. They are collected on Jev AI cases, which updates when the daily intake adds a post. A repository name in a post is a pointer. Confirm it on the projects page before you depend on it. OpenRouter traffic labels in the fact card, including sel-jev-rubric-judge, Waura, mirasim, JeVinci, clara jev triage, are app names. They are not automatically GitHub URLs.

Is the hosted model open source?

No source in this handbook shows TypeSafe AI releasing the Jev model weights. Official GitHub entries we index are clients, adapters, and agent skills, each with its own license. Community projects grouped as open models say they are independent and do not reproduce TypeSafe’s private weights. Treat the hosted Jev API as a closed model unless TypeSafe publishes weights, and read the license on any repository before you reuse it.

Jev AI vs LLM covers when a chat model is the wrong tool, what “Jev alternative” can honestly mean, and how Jev compares with GLiNER. Several open-model repositories in the GitHub directory describe themselves as independent. Read those READMEs. Do not treat a similar JSON shape as TypeSafe’s weights.

How this unofficial handbook is organized

Jev Notes does not sell API access, does not host the model, and does not claim a partnership. The pages exist so a developer can answer one question without reading the whole docs site:

• What is Jev AI — the decision model, noul, choice, score, and when not to use it.
• Primitives — field rules and a request builder.
• API quickstart — the direct System One call.
• Jev on OpenRouter — recorded model ids and the Decisions API warning.
• Jev as a judge — rubric, eval, and score posts already in the case catalog.
• Pricing calculator — daily and monthly cost at the published input rate.
• Jev vs LLM — prose versus a typed decision, plus alternatives and GLiNER.
• GitHub projects and skills — public repositories.
• Cases — community posts, updated when the daily intake adds one.

Primary sources remain typesafe.ai, docs.typesafe.ai, docs.typesafe.ai/llms.txt, and console.typesafe.ai. If this page and the docs disagree, the docs win. The launch post linked from the channel map is Introducing System One Models and Jev.

What this page will not invent

Headcount, funding, and a headquarters are not in this repo’s fact card, so they are not stated here. A model id, a price, or a latency number appears only when it is already in the fact card or the channel map. Example JSON elsewhere on the site is illustrative unless it is labeled as a measured run. The founder line is only what public materials used here already say: Diogo Almeida, described as an OpenAI alumnus.

FAQ

Is Jev AI open source?

No source in this handbook shows TypeSafe AI releasing the Jev model weights. Official GitHub entries we index are clients, adapters, and agent skills, each with its own license. Community projects grouped as open models say they are independent and do not reproduce TypeSafe’s private weights. Treat the hosted Jev API as a closed model unless TypeSafe publishes weights, and read the license on any repository before you reuse it.

Is TypeSafe AI the same thing as Jev AI?

No. TypeSafe AI is the company. Jev AI (System One) is their decision model. This handbook is not affiliated with TypeSafe AI.