TypeSafe AI
Jev
A decision API for structured judgment
- Model
- Jev 1.13
- Model ID
- jev-1.13.0
- Context
- 64k tokens
- Input price
- $0.042 / million
- Output tokens
- free
- Latency
- 70ms-500ms
- Throughput
- 250,000 tok/s
- Request cap
- 1,200 / min
Limits can change without notice.
Jev is TypeSafe AI's flagship model and the first System One model. You send a state plus typed questions; it returns Choice, Score, or Noul answers with probabilities instead of a chat transcript. The live build is Jev 1.13 (jev-1.13.0). Input costs $0.042 per million tokens, output tokens are free, and TypeSafe quotes 70ms-500ms end-to-end with most queries around 100ms. English is the strongest language. Open Primitives before you write a request, or read the Jev 1.13 card for aliases, context, and rate limits. Limits can change without notice during early access. Pin jev-1.13.0 if you have already tuned those gates.
What Jev is
A System One model is built for software that has to act. Jev reads natural-language state — a string, a JSON object, or an array of text — and answers a map of questions in one POST to /v1/systemone. It does not write replies, produce code, or explain its reasoning. Images, audio, and video stay out of the request; preprocess those into text if you need them.
The three question types are Choice (one option from up to 255), Score (an ordered scale of 2 to 10 levels, which can land between two levels), and Noul (P(yes) in [0, 1] with no separate confidence field). Questions in one call share state, run in parallel, and barely change latency as you add more. Question IDs exist for your code and are not sent to the underlying model.
Jev is named after economist William Stanley Jevons. System One borrows Kahneman's fast System 1: a quick judgment, then your code does the slow work. TypeSafe trains with Reinforcement Learning for Calibrated Decisions (RLCD), not RLHF chat tuning. Schema matching is guaranteed, so the response cannot emit a free-form type error.
TypeSafe AI and the launch
TypeSafe AI is made in San Francisco. Diogo Almeida is founder and CEO and a co-inventor of RLHF. LinkedIn lists the company as privately held with 11-50 employees. Jev opened in early access on 15 September 2026 after about two years in stealth. The launch post describes a new architecture, a parallel sampler, and RLCD as the training method.
Customer requests and responses are not used to train Jev. The same weights serve every account; you shape answers with state, instructions, and criteria rather than a LoRA. Higher rate limits exist on custom and enterprise plans via sales@typesafe.ai. The playground and API keys live on console.typesafe.ai after login.
Official launch video
Jev opened in early access
TypeSafe spent about two years in stealth before this date. No other public Jev version sits between stealth research and Jev 1.13.
Community directory and press
Awesome Jev indexes community clients, integrations, games, and articles and is not affiliated with TypeSafe AI. Treat those packages as community work unless the README points at github.com/typesafe-ai. Official Python is typesafe-sdk; official JavaScript is @typesafe-ai/sdk.
Cloudflare Workers AI lists the model as typesafe/jev. Steam store search for Jev TypeSafe returned no apps: Jev is an HTTP API, not a store title. Discord is discord.gg/typesafe. GitHub org typesafe-ai publishes the JS SDK and related skills. TypeSafe's YouTube channel exists as a brand page without uploads, so this wiki uses the official homepage video instead of a YouTube facade.
A first request, then a gate
The shortest path into Jev is the playground at console.typesafe.ai/playground: paste a ticket, add a Noul for urgency, then a Choice for the owning team. That same map becomes a curl to POST /v1/systemone or a TypeSafeClient.system_one call. Pin jev-1.13.0 once you like the answers; follow jev-latest only if you want alias drift.
Before any automatic side effect, read confidence on Choice and Score and treat a Noul near 0.5 as uncertainty. Scan the nine jaggedness modes for the question you just wrote. If you asked Jev to count, compare dates, or draft a sentence, move that work out. The primitives page is the compare; the API page has the install lines; jaggedness is the filter you run before shipping.
Price the loop against $0.042 per million input tokens and free output. The Doom demo's about $7/hour at 10 queries per second is the public illustration of a tight loop, not a bill you must hit. Rate limits of 250,000 tokens per second and 1,200 requests per minute can move without notice while early access scales. Sales can raise them on custom plans.
Keep this homepage as the punch-card index: live facts first, then what Jev is, then who shipped it, then the official video. From here the next clicks are Primitives for the compare, Models for the card, Jaggedness for the nine failure modes, and API for the install lines. Discord and GitHub are linked from the community section when you want unofficial clients.
Jev opened in early access on 15 September 2026. The facts on this page — 64k context, $0.042 per million input tokens, free output, 70ms-500ms, jev-1.13.0 — are the live card TypeSafe published with that launch. When a newer versioned ID ships, the model page should show it; this homepage should keep pointing at the current card rather than at a frozen screenshot.
If you only remember three clicks from this homepage, make them Primitives, Jaggedness, and API. That is the loop TypeSafe documents: pick a question type, avoid the known failure modes, then call POST /v1/systemone with jev-1.13.0 or jev-latest.
TypeSafe is made in San Francisco; Diogo Almeida is CEO. Those company facts sit under the launch heading, not in the live ledger.
Jev release updates
New Jev versions, benchmark changes and API notes — occasional email, no tracking pixels.
We store only your address and the date. Unsubscribe any time by replying to any message.
Sources