Jev Pricing and Access, Explained: What One Decision Costs, How to Get In, and Why the Model Is Now the Cheapest Line

TL;DR: Jev, the "System One" decision model TypeSafe AI launched on 15 September 2026, costs $0.042 per million input tokens, and its output is free because it returns typed answers, not text. A support-ticket-sized decision — around 300 input tokens — costs about $0.00001; TypeSafe's own heavier workflow cases average $0.0004. On TypeSafe's published evals that is roughly 8x cheaper than the cheapest LLM tested and around 440x cheaper than the most expensive, at accuracy in the same band as the mid-tier LLMs. The practical consequence for anyone running automations: the model stops being the line that matters. What you pay for a routed ticket or a scored lead is now the platform run, the LLM steps that remain (reading documents, writing replies) and — above all — the minutes a person spends on the cases the model was unsure about. Price the review queue, not the tokens.

Buildable version: the support ticket deflection workflow — every ticket is classified and routed on a confidence score before anything is answered, which is exactly the step a decision model prices at fractions of a cent.

What Jev costs: the price list

Jev pricing has two lines, and one of them is zero. TypeSafe charges only for what goes in — the text you ask about (the "state") plus the questions — and nothing for what comes out, because what comes out is a label, a score or a probability rather than generated text.

The LLM column is TypeSafe's own framing from its launch post, not a price list of any one provider; the Jev column is from its models page. The rate limits carry a warning on that page that they can change without notice while TypeSafe adds capacity — worth knowing before you put a month-end batch on it.

What one decision costs, worked out

The arithmetic is short enough to do on a napkin: cost per decision = (tokens in the state + tokens in the questions) × $0.042 ÷ 1,000,000.

TypeSafe's quickstart example — a three-sentence support message asking whether it is urgent — reports 296 input tokens. That is $0.0000124 per decision, or about $0.12 for 10,000 of them. Asking more questions of the same text barely moves it: the state is read once and each extra question adds a few dozen tokens, so a ticket asked for category, urgency, sentiment and "does this need a person" in one call costs about the same as asking one.

Real workflow cases are bigger. TypeSafe's published workflow evals feed each case a full record — an alert with the machine's history, an invoice with its purchase order and delivery note — and ask many questions per case. Those average $0.0004 per case, ranging from $0.0001 (security alerts, customer service) to $0.0011 (invoice processing, where the documents are long).

Measure it yourself: every response reports its input token count. Run fifty of your own real cases through, take the median token count, and multiply. Do not estimate from the quickstart — your records are longer than a three-sentence email.

Jev vs an LLM on the same decision

The fair comparison is the same workflow, the same questions, different model. That is what TypeSafe's evals measure: four workflows (security incidents, agent-run review, invoice processing, customer service), with every model asked the same decomposed questions and scored against the averaged answers of two frontier models run at high reasoning.

Read it in two directions. Against the priciest configurations, the gap is enormous: the "444.6x cheaper, 193.6x faster" headline on TypeSafe's homepage lines up with Opus 5's cost and Sonnet 5's time in this table. Against the cheapest LLM tested (Luna), Jev is about 8x cheaper and 30x faster at similar accuracy — still a large difference, but not the one on the billboard. And the two most accurate models, Sol and Opus 5, are five to six points ahead; on invoices alone the best LLM scores 79.1% to Jev's 61.8%.

Independent measurements land well below the billboard, and they shrink as the measurement widens from one call to a whole run:

The last row is the one to keep in mind when you plan a workflow: the model is one step among several — fetching the record, the platform run, the LLM steps that remain — and the end-to-end speed-up can only be as large as the model's share of the time. The same holds for cost, which the next section is about.

TypeSafe lists the caveats itself: the workflows were written by its own team, the reference answers come from OpenAI and Anthropic models, the LLMs ran through TypeSafe's wrapper (which makes them slower and dearer than calling them plainly), and timing was measured from laptops on the US West Coast. Treat the table as direction, not as a quote for your workload. The comparison of which workflow steps each model should handle goes through where the accuracy gap matters.

Where the money goes once the model is nearly free

Here is the part the token calculators miss. When a decision costs a hundredth of a cent, it is no longer the expensive part of the step it sits in. For a routed support ticket or a scored lead, the cost stack looks like this:

The review line is the one that dominates, and it is the one the model's price has nothing to do with. An illustrative case: 3,000 tickets a month, 10% below the confidence threshold, two minutes each — that is 10 hours of someone's month, which costs more than every other line in the table combined. Move the threshold so 5% go to review and you halve it; move it so 2% go and you had better be sure the model is right on the other 98%.

That is why the useful question about Jev is not "how cheap" but "how often is it sure, and is it right when it is". The shadow test for Jev's accuracy is how you find your own answer before any of this goes live.

Measure it yourself: for any automated decision you already run, count last month's cases that a person touched and multiply by a realistic two minutes. Compare that to the model bill. For most teams the first number is already the bigger one.

Will the price last?

Nobody knows yet, including TypeSafe. Its launch post says plainly that it cannot prove the price is not subsidised and will need time to show it is sustainable — while also saying it expects prices to go down, not up. Three sensible precautions follow:

• Keep the step swappable. A decision step that asks "which team, how urgent, does it need a person" can be answered by Jev, by a small LLM or by a larger one. Design it so changing the model is a configuration change, and keep an LLM path you have actually tested. Since 1 October that has become easier: Cloudflare's Clef accepts the same requests as Jev and its weights are free to download, and OpenAI and AWS have released decision models of their own — the alternatives are listed in the accuracy guide.
• Pin the version. TypeSafe's jev-latest name moves when a new release ships. If you have tuned thresholds against one version, pin that version and move deliberately.
• Mind the data terms. Zero data retention is offered to enterprise customers, and TypeSafe says its service is currently based on the US West Coast. For personal or regulated data, that decides more than the price does.

How to get access: no waitlist, and five ways in

There is no Jev waitlist any more. TypeSafe launched Jev in early access on 15 September with a waitlist, and on 20 September opened it to everyone: you sign up in its console, where access, keys and a playground live. Four platforms you may already pay for offer it too, each with its own account and billing:

Which door you use is a billing and data-terms decision more than a technical one: the price is the same wherever one is listed, and the answers come from the same model.

For a business that wants the result rather than the access — tickets routed, leads scored, spend classified — the waitlist that matters is a different one: whoever runs your workflows has to put Jev on a step, test it against what that step uses today, and own the thresholds. That is the one below.

What it costs to run as a workflow

A routing or scoring step never runs alone — it sits in a workflow that collects the ticket or the lead, asks the questions, acts on the answer and queues the unsure cases for a person. The support ticket deflection workflow is a free template if you want to build it yourself, or $297/month to have it run for you (up to 3,000 tickets, one help desk, one knowledge base); multi-language or actions inside your product are a custom build from $4,000. At that volume the model is the smallest line on the invoice whichever model answers — which is the point. Choose the model on accuracy and on how honest its confidence is, not on the price per token.

Questions, answered

How much does Jev cost?

$0.042 per million input tokens, and output is free. A short support message costs about $0.00001 to classify; TypeSafe's heavier workflow cases average $0.0004. At 10,000 decisions a month that is between about $0.12 and $4, depending on how much text each decision reads.

Why are Jev's output tokens free?

Because Jev does not generate text. It returns a choice from options you defined, a score on a scale, or a yes/no probability, so there is almost nothing to produce. TypeSafe calls the output "too cheap to meter" and charges only for the text going in.

Is Jev cheaper than Claude or GPT?

Much cheaper per decision: on TypeSafe's own evals, about 8x cheaper than the cheapest LLM tested and around 440x cheaper than Opus 5, at accuracy similar to the mid-tier models and five to six points below the best. It cannot replace an LLM for anything that writes, reads images or scans, or calculates — those steps stay on an LLM or in plain rules.

Is there a Jev waitlist?

Not any more. TypeSafe launched Jev in early access with a waitlist on 15 September 2026 and opened it to everyone on 20 September; you sign up in TypeSafe's console. It is also available through Vercel's AI Gateway, OpenRouter, Cloudflare's AI Gateway and Netlify's AI Gateway, on their own accounts and billing.

Is Jev pricing subsidised?

TypeSafe says it cannot prove it is not, and that it expects prices to fall rather than rise. Until there is a longer track record, keep the decision step swappable and a tested LLM fallback in place.

Sources and further reading

• Introducing System One Models & Jev — TypeSafe AI (launch post, 15 Sep 2026)
• Models: price, rate limits, context length — TypeSafe docs
• API reference and quickstart usage figures — TypeSafe docs
• Workflow evals: accuracy, cost and time per case — TypeSafe
• Legal: data processing and zero data retention — TypeSafe docs
• Jev is now available to everyone, no waitlist — TypeSafe AI on X (20 Sep 2026)
• Connecting through a gateway (OpenRouter, Vercel AI Gateway) — TypeSafe SDK docs
• TypeSafe AI's Jev now available on AI Gateway — Vercel (16 Sep 2026)
• Jev 1.13 — OpenRouter
• Jev — Cloudflare AI docs
• TypeSafe Jev now available in AI Gateway — Netlify (7 Oct 2026)
• Mini vibe check: TypeSafe's Jev judged everything I've written in 0.7 seconds — Mike Taylor, Every (15 Sep 2026)
• Jev in fx's safety reviewer: up to 18x faster at p95 — Guillermo Rauch, Vercel, on X (16 Sep 2026)
• Jev vs GPT-6 Astra in a search-and-read loop — keiranhaax on X (17 Sep 2026)
• Support ticket deflection workflow: blueprint, template and price
• AI agent cost optimization