Why Jev Is So Cheap: Who Built TypeSafe AI, and What It Cannot Do

Jev is not a cheaper ChatGPT. It is TypeSafe AI’s first System One model: software sends it state plus a fixed list of answers, and it returns typed choices with calibrated probabilities. It does not write text. Official list, checked 20 September 2026: $0.042 per million input tokens, output free. That is 238× cheaper on input than Claude Fable 5.1 at $10 per million. The first-party API is still early access / waitlist at typesafe.ai.

If you need a chatbot, a coding agent, or a summary, stop here and use the wizard. Jev is for routing, scoring and guardrails inside code. The rest of this page is who built it, why the rate is that low, and the trade-offs.

What Jev actually is

TypeSafe calls this class System One: intuition for software, not a conversation. You send the model a state (the facts the program already has) and one or more typed questions — classify this ticket, pick a route, score a lead, say whether this action is allowed. Jev answers with labels and probabilities. TypeSafe says those scores are calibrated, so you can set a threshold: act automatically above 0.9, escalate below.

It will not draft an email, write a function, or explain itself in paragraphs. TypeSafe is explicit: Jev gives up string generation on purpose. That is the product, not a missing feature. The homepage also lists what it will not do: vision, math, dates. Maximum 255 choices per question. Hosted only — no published weights.

The name is a nod to William Stanley Jevons. TypeSafe’s bet is the Jevons Paradox: when intelligence gets an order of magnitude cheaper, people do not spend less on it. They find more places to spend it. Jev is the first public model they built for that thesis.

Who created it

TypeSafe AI is a San Francisco lab founded in 2024. It came out of stealth on 15 September 2026 with about $40 million in seed funding led by DCVC. The three founders:

• Diogo Almeida, CEO. Former OpenAI researcher. TypeSafe and the launch coverage describe him as a co-inventor of RLHF / ChatGPT. He left OpenAI about two years before the Jev launch to start the company. His public writing is about how models learn — distillation, teacher probabilities, copying tokens versus imitating answers. Jev is trained, he says, on synthetic data with a method he calls reinforcement learning from calibrated decisions (RLCD).
• Erik Gafni, CTO. Repeat founder (Ravel, multimodal AI for DNA sequencing). Early employee at Invitae and Freenome. TypeSafe’s team page says he specialises in production AI systems.
• Sasha Sheng, COO. Ex-research engineer at Meta / FAIR (News Feed, AI Experiences, AI Research). Published at NeurIPS and ECCV. Builder / hackathon organiser more than a research-only profile.

The lab also went through the 2024 AWS Generative AI Accelerator. Architecture details are thin. TechCrunch noted that outsiders suspect a closed-weight stack on top of an existing LLM; Almeida has not published the weights or a paper that would settle that. Treat Jev as a hosted API with a new interface, not as an open model you can inspect.

Why it is so cheap

The list price is $42 per billion input tokens. That is $0.042 per million. Output is free because there is almost no output: a handful of labels and floats, not a 800-token essay. Four reasons the number can be that low — only the first three are engineering. The fourth is money.

1. It does not generate strings. Autoregressive decoding is the expensive half of a normal LLM call. Jev scores a fixed answer set. TypeSafe says it can score hundreds of those answers in one pass. You are not paying for tokens the model invents.

2. The job is narrower. Classify / route / score is cheaper to serve than “be a general assistant.” There is no tool-calling loop, no 128K essay, no vision encoder. 64k context is the cap. That is enough for a ticket plus policy text. It is not enough for a repo.

3. Latency is the other product. TypeSafe quotes 70–500 ms end-to-end from the US West Coast, and claims on the order of 100× to ~194× faster than frontier chat models on its own System One workflows. Fast inference is cheap inference if you skip generation. The 444.6× cheaper / 193.6× faster figures on the homepage are TypeSafe’s high end on their evals. Do not paste them into a board deck as independent benches.

4. TypeSafe cannot prove the price is not subsidized. That sentence is theirs, on the site. $40M of seed and a waitlist launch is exactly when a lab prices for adoption, not for GPU recovery. Compare it to DeepSeek V4.1 Flash at $0.15 / $0.60 off-peak: Flash is still a chat-and-vision model. Jev undercuts it on a token bill because Jev is not in that market. The rate can move the day the waitlist opens or the seed is spent. Date it: 20 September 2026.

A real monthly bill

Same workload we use on the Flash page: 25,000 requests a month, about 1,500 input tokens and 800 output tokens each. That is 37.5 million input tokens and 20 million output tokens. For Jev the 800-token “output” does not exist — you still send the 1,500-token state. Dated September 2026.

Jev: 37.5 × $0.042 = $1.575. Flash off-peak miss: 37.5 × $0.15 + 20 × $0.60 = $17.63. Fable: 37.5 × $10 + 20 × $50 = $1,375. Run the mix in the token calculator — Jev is now a row under TypeSafe AI.

That table is honest only if you accept the job change. A $1.58 Jev bill does not replace an $18 Flash bill. It replaces the classify/route step you were overpaying a chat model to do. The rest of the pipeline — the email, the code, the summary — still needs a generator. That is the same shape as the 2026 API price war, except the cut here is architectural, not geographic.

Advantages

• Price. $0.042/1M input and free output is the cheapest row we track for a hosted intelligence API, as of 20 September 2026. Fine to call at every form submit, every agent step, every moderation check.
• Speed. Tens to hundreds of milliseconds, not seconds. Usable in a request path, not only in a batch job.
• Typed answers. You get a label and a probability, not a paragraph you have to parse. Software can branch on it. TypeSafe’s claim is that Jev cannot hallucinate a string because it never writes one.
• Calibrated confidence. If the scores are as advertised, you can set “auto vs human” without a second model. That is the whole point of RLCD.
• Parallel choices. Hundreds of options in one call. A router with 40 destinations is one request, not 40.
• US host, no China hop. Different risk profile from DeepSeek or Kimi. Still a cloud API — prompts leave your network — but the jurisdiction is San Francisco, not Beijing.
• Hooked into agent stacks already. Official API, a LangChain package (langchain-typesafe), a Claude Code / Codex skill, and community OpenClaw plugins. Not a chat-completions clone, but you are not starting from a raw HTTP client either.

Disadvantages

• It is not a general AI. No chat, no code, no summaries, no images, no math, no dates. If your product is “talk to the model,” this is the wrong SKU.
• Waitlist. First-party access is early access. Third-party aggregators may sell a key; that is not TypeSafe’s SLA. Plan for “we applied” as the current state.
• Closed weights, hosted only. You cannot self-host, inspect, or air-gap it. Almeida has not published an architecture paper that would let you reproduce the model.
• The price may be a promo. TypeSafe said so. Seed-funded launch rates move. Do not write a three-year cost model on $0.042.
• 64k context, 255 choices. Fine for a ticket. Not fine for a codebase or an open-ended taxonomy.
• New interface. You reshape the prompt into state + typed questions. Teams that only know Chat Completions will feel the rewrite.
• Quality is TypeSafe’s number. They put Jev at 67.8% on their System One workflow eval, level with GPT-5.6 Terra on that same harness. We have not rerun it. Prove it on ten of your own decisions before you delete the Claude call.
• No consumer app. There is nothing to “try in the browser” the way there is for ChatGPT or DeepSeek. If you are picking a daily driver, start at the wizard.

Jev is cheap because it refuses to be a chatbot. $0.042/1M input, output free, 70–500 ms — for classify, route and score. Everything else still needs a generator.

If the job is “pick a model for this month’s budget,” use the wizard. If the job is “price a million classify calls,” use the token calculator and select Jev (TypeSafe). Prices on this page are TypeSafe’s public list as of 20 September 2026.