Jev AI Pricing Explained: $42 Per Billion Tokens, Free Output
Jev prices input tokens at $42 per billion and gives output tokens away free. Here's what that pricing model means for real-time AI apps.
What is Jev’s pricing model?
Jev, the “system one” foundation model from Typesafe AI, prices input tokens at $42 per billion tokens and gives output tokens away for free. One independent test measured the rate slightly differently, at 4.2 cents per million input tokens, which lines up with the same per-billion figure. Either way, the headline is the same: input is cheap, and output costs nothing because Typesafe says its output tokens are “too cheap to meter.” That pricing structure only makes sense once you understand what Jev is actually built to do.
TL;DR
- Jev charges only for input tokens, at a published rate of $42 per billion tokens (about 4.2 cents per million), with output tokens priced at zero.
- The model isn’t a chatbot, it returns structured decisions (a choice from options, a score, or a probability) instead of generated prose, which is why free output is economically viable.
- Real-world tests showed extremely low costs, including a Minecraft bot session that ran for two minutes on roughly 150,000 tokens for about a cent, and a drone-navigation demo built in 15 minutes for around 10 cents.
- Speed matters as much as price, with reported model evaluation times in the 92 to 214 millisecond range in one independent playground test, far faster than typical LLM generation.
- The free-output claim held up in testing, since Jev’s answers are short structured decisions rather than long text, so there’s little to meter in the first place.
- Total application cost is still higher than the sticker price, because you need other components (data prep, validation, sometimes a separate text-generating model) to build a full system.
- The pricing model favors high-frequency, low-latency use cases, like game AI, robotics control loops, and support-ticket routing, over tasks that need long-form generation.
Why is Jev’s output free?
Jev doesn’t write sentences. It answers one of three question types: a choice (pick from a list of options you supply), a score (rate something on a scale you define), or a null question (return a probability that a yes/no statement is true). Every response is a small structured value, not a paragraph.
That’s the mechanical reason free output is possible. A standard LLM generates output token by token, and each token costs compute and money. Jev’s outputs are so short and structured that Typesafe apparently decided metering them wasn’t worth doing. In one independent test covering eight separate playground requests, the combined input came to 4,148 tokens, and at the published input rate that added up to a fraction of a cent in charges, with no separate output charge at all.
This is a fundamentally different cost shape than what developers are used to with GPT-style models, where output tokens are often priced higher than input tokens because generation is the expensive part. Jev inverts that by removing generation from the picture almost entirely.
How does the cost-per-token economics compare to standard LLMs?
Typesafe’s launch materials describe Jev as roughly 100 times cheaper and 100 times faster than typical large language models for the kinds of tasks it targets. Independent demos back up the general shape of that claim, even if the exact multipliers weren’t independently reproduced:
- A Minecraft-playing bot reportedly ran for about two minutes, consuming roughly 150,000 tokens, for a total cost of about one cent.
- A self-driving-style simulation built in under an hour reportedly ran at near-zero marginal cost per decision cycle.
- A drone navigation demo built in about 15 minutes cost roughly 10 cents to run.
- A browser automation demo (built by a third-party developer combining Jev with browser-use tooling) completed a flight search task in about 7 seconds for a reported cost of $0.0039, under half a cent, based on logged model usage (excluding surrounding infrastructure costs).
Compare that to a typical LLM agent loop, where each decision requires a full prompt, a generated explanation, and often a chain-of-thought before an action is chosen. That approach is slow and token-hungry precisely because the model is optimized to produce fluent language, not fast decisions. Jev skips the explanation step and returns a decision directly, which is where both the speed and cost advantage come from.
Is Jev worth it for real-time applications?
For workloads that need many small decisions made quickly and cheaply, the pricing model looks compelling on paper. Game AI, robotics control loops, browser automation, and support-ticket triage all involve repeated structured judgments rather than long creative output, which is exactly what Jev is priced and built for.
But cheap tokens don’t automatically mean a cheap or reliable system. Independent testing found real limitations worth flagging:
- When a question’s answer options didn’t actually cover the situation (asking about a cafeteria’s hours but only offering billing, technical, or sales as choices), Jev still picked one of the offered categories, with a low confidence score, rather than signaling that none of the options applied.
- Confidence values need their own validation. A response with 90% confidence isn’t guaranteed to be correct 90% of the time; that has to be tested against your own data.
- The candidate list matters as much as the model. In an extraction test, Jev correctly picked the right value from a document, but only because the correct value was already present in the list of choices it was given. If the input pipeline misses the right candidate, Jev can’t invent it.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
So the honest answer is that Jev’s low, predictable cost per token is real, but the total cost of a working application still depends on the code around it: the logic that prepares input, the fallback paths for “none of these fit,” and any human review step for edge cases.
What tasks fit Jev’s pricing model best?
Based on the demos and independent tests circulating so far, Jev’s economics make the most sense for:
- Routing and classification, such as sorting support tickets by department, urgency, or sentiment, where one test showed it correctly separating a billing issue from a false urgency signal.
- Fast-loop control tasks, like game characters, drone navigation, or simulated driving, where a decision has to be made every fraction of a second and a full LLM response would be too slow and too costly.
- Claim verification against evidence, such as checking whether an agent’s reported outcome (“draft saved successfully”) matches the actual tool output (“permission denied”), which one test showed Jev flagging correctly with a 93% probability of an unsupported claim.
- Hybrid agent pipelines, where Jev handles the “what to do next” decision and a separate small language model handles any text that actually needs to be generated, as demonstrated in a browser automation project that paired Jev with another model for text input.
It’s a poor fit for anything that needs long-form writing, reasoning chains, or creative generation, since that’s not what the architecture or the pricing model is built around.
Frequently Asked Questions
How much does Jev cost per token?
Jev’s published rate is $42 per billion input tokens, which works out to about 4.2 cents per million tokens. Output tokens are free.
Why are Jev’s output tokens free?
Jev returns short structured answers (a selected choice, a score, or a probability) instead of generated text, so there’s very little output to meter compared to a standard chatbot-style model.
Does free output mean an app built on Jev is free to run?
No. You still pay for input tokens, and you likely need other infrastructure, such as code to prepare inputs, validate outputs, or a separate model for any actual text generation, all of which add cost beyond Jev’s own pricing.
Is Jev cheaper than models like GPT-style LLMs for every task?
Only for the tasks it’s designed for: fast, structured decisions rather than open-ended text generation. For long-form writing or complex reasoning, a general-purpose LLM is still the more appropriate tool, and Jev’s pricing advantage doesn’t apply the same way.
Can I trust Jev’s confidence scores as accuracy guarantees?
Not without testing. Independent evaluation found that restricting Jev’s answer options can prevent it from inventing new categories, but it can still select a poor-fit answer from the options given, sometimes with a low but non-trivial confidence score. Confidence values should be validated against your own data before being used to gate real decisions.
