The Jev Build Report

How many builds, and how fast did they arrive?

239 builds in 6 days. The curve is not the model’s, it is ours: the directory opened on 18 September, which is why that day carries the backlog. The steady rate since, around fifty a day, is the better number to quote.

Half of them are code. 121 builds (50.6%) are public GitHub repositories and 95 (39.7%) are posts on X showing something running. The rest are agent skills, live sites and articles.

What are people pointing Jev at?

Small tools, mostly. 77 of the 239 builds (32.2%) are plugins, extensions and apps with a Jev call inside. Agents and browsers come second with 52 (21.8%). The two categories a launch-week reader might expect to lead, trading and robotics, account for 4 and 3 builds.

The split by form runs across the categories in a way worth noting. Tools and apps are mostly repositories, 46 against 19 posts. Content and growth is the opposite, 22 posts against 7 repositories. Where the work is a product, people ship code. Where the work is a workflow, they ship a screenshot.

What does a decision actually cost?

The median published figure is $0.000068 per decision, which is about 14,727 decisions for a dollar. That comes from the 15 builds whose authors published both a volume and a total cost, so the division is ours and the two inputs are theirs. The spread across those 15 builds is wide: from 1,429 decisions per dollar at one end to 170,900 at the other, a factor of about 120. What sits between them is prompt length, not model price: Jev bills $0.042 per million input tokens for everyone, and output is free.

One row is an upper bound: Every reported its 1,709 editorial judgments as costing under a cent, not as a total. It is marked with a < and kept in the median, where it sits at the cheap end either way.

The practical reading is not that Jev is cheap. It is that the queue in front of it stops mattering. A build that scores 100,000 posts for 67 cents has no reason to sample, and several of the builds in the table are things nobody would have written a budget for a month ago.

How long does one decision take?

A median of 300 ms across 19 builds that published a per-decision time. 18 of the 19 are under one second and 16 are at or under half a second. The fastest published figure is 50 ms, for a form fill; the slowest is 1,000 ms, for a call carrying 61 questions at once. TypeSafe’s own published band is 70 to 500 ms, so the builds land on it from the slow side, which is what network and application code around a model normally costs.

Batch throughput is the same fact seen from above. Across the 13 builds that published both a volume and a run time, the median is 22.5 decisions a second, and the fastest reported run is 4,902 a second.

Who is building: TypeScript, not Python

Of the 126 repositories in the directory that were created on or after launch day, 124 have a primary language on GitHub. 72 of them (58.1%) are TypeScript or JavaScript and 38 (30.6%) are Python. A model released with a Python SDK and a research paper’s vocabulary was picked up first by people who build web applications.

Those 126 repositories hold 17,556 stars between them, and the distribution is as skewed as it usually is on GitHub: the five largest hold 74.6% of the total, the median repository has 4 stars, 74 have fewer than ten, and 98 have fewer than fifty. Four are over a thousand.

The people are spread just as thin. The 239 builds come from 216 different authors, and 198 of them have published exactly one. Only 18 have published two or more. Week one has a crowd, not a community.

What this report cannot tell you

The gaps matter as much as the figures, so here they are in full.

• The figures are author-reported. Nothing on this site is benchmarked by us. We check that a claim is published at a URL anyone can open, not that it reproduces. Only 102 of 239 builds publish any figure, and 35 publish a cost.
• Accuracy is mostly missing. Cost and speed are easy to post; a held-out set is not. Any comparison between Jev and a larger model that uses only the numbers on this page is a comparison of price and latency, not of quality.
• The directory is curated, not exhaustive. It is what one person found and checked. It undercounts private and internal work by definition, and it undercounts anything published somewhere we do not read.
• Stars are a snapshot. They were read on 2026-09-20 and they move daily. Quote them with the date.
• Six days is not a trend. Nothing here supports a projection.

The outside numbers we checked the same day

For context, and so nothing on this page rests on a figure we have not opened ourselves on 2026-09-20.

Hacker News launch thread: 1,915 points · 501 comments

Posted 15 Sep 2026, 19:25 UTC. Counts via the HN Algolia API.

Published input price: $0.042 per million tokens, output free

TypeSafe’s launch post. It is the price every cost on this page was paid at.

TypeSafe’s own headline: 193.6× faster, 444.6× cheaper

The vendor’s figure, from its own workflow evals. Not ours.

Most-starred Jev repository: browser-use/jev-ultrafast, 9,134 stars

Created 16 Sep 2026. Star count via the GitHub API.

Method

1. The count. Every build in this report is an entry in the Made with Jev directory, each with a public URL. Counts by day, category and form are computed from that list when the page is built, so they match the directory exactly and cannot drift from it.
2. Cost per decision. Taken from the 15 builds whose author published both a volume and a total cost. Cost divided by volume. Rows where the author gave a bound rather than a total are marked. Nothing is converted, inferred from token counts or estimated from the price list.
3. Decision time. Taken from the 19 builds that published a per-decision time. Where the author gave a median we use the median; where they gave a rate, such as ten queries a second, we use its reciprocal, and the table says which.
4. Throughput. Taken from the 13 builds that published both a volume and a wall-clock run time, divided.
5. GitHub. Every github.com URL in the directory was resolved once through the GitHub GraphQL API on 2026-09-20. 131 of them still resolve; 126 were created on or after 2026-09-15 and are the cohort used for every star and language figure here. The five that predate the launch are excluded from those figures, so stars a project earned before Jev existed are not read as launch-week interest.
6. Medians, not means. Every distribution on this page is long-tailed. A mean would be a number no build resembles.

The underlying rows are published as JSON at /jev-statistics.json, including the full GitHub snapshot, so any figure above can be recomputed.

Facts you can quote, already written

Each line below is a complete sentence with its figure, its scope and its date in it, so it can go into a piece as it stands. The numbers are the ones computed above, not rounded again for effect.

• 239 public projects built with Jev were published in the 6 days after its release, according to the Made with Jev directory (measured 20 September 2026).
• Across the 15 Jev projects that published both a volume and a total cost, the median works out at $0.000068 per decision — about 14,727 decisions for a dollar.
• The median published time for a single Jev decision is 300 ms, and 18 of the 19 published figures are under one second.
• 58.1% of the 126 Jev repositories created since launch day are written in TypeScript or JavaScript, against 30.6% in Python.
• Those 126 repositories hold 17,556 GitHub stars between them, but the median repository has 4 and the five largest hold 74.6% of the total.
• The 239 builds come from 216 different authors, 198 of whom have published exactly one.
• Only 102 of the 239 public Jev builds publish any figure for cost, speed or volume, and 35 publish a cost.

Writing something that needs a person rather than a number? Jon Kraayenjon compiles this report and is available for comment on what builders are shipping with Jev, what the published costs do and do not show, and how the directory is put together. Ask on X, or say which figure you need and it will be cut from the data and sent back.

How to cite this

Free to quote, chart or reprint, in any outlet, with a link to this page. No permission needed, no paywall to clear and no embargo. The report is re-measured as the directory grows, so a citation that carries the date stays correct.

Common questions

How many public Jev builds are there?

239 as of 20 September 2026, six days after Jev was released. Made with Jev counts a build once its author has published it somewhere anyone can open: a repository, a post, a live site or an article. 102 of them publish at least one figure for cost, speed or volume.

What does one Jev decision cost in practice?

The median published figure is $0.000068 per decision, or about 14,727 decisions per dollar, across 15 builds whose authors gave both a volume and a total cost. The range runs from 1,429 decisions per dollar to 170,900.

How fast is Jev in published builds?

The median published time for one decision is 300 ms across 19 builds. 18 of the 19 are under one second, and 16 are at or under half a second. The fastest is 50 ms, the slowest 1,000 ms.

What language are Jev projects written in?

Of the 126 repositories created since launch day, 124 have a primary language on GitHub. 72 of those are TypeScript or JavaScript (58.1%) and 38 are Python (30.6%).

How many people are building with Jev?

The 239 public builds counted here come from 216 different authors, 198 of whom have published exactly one build and 18 of whom have published two or more. That counts public work only: private and internal projects are invisible to this or any other count.

Is Jev cheaper than using an LLM for the same job?

On price, in every published side-by-side, by a wide margin: the median published Jev figure is $0.000068 per decision. On quality, the public record does not say. Cost and speed are easy to post and a held-out accuracy set is not, so only 102 of 239 builds publish any figure at all. Treat a cost comparison as a cost comparison.

Where can a journalist get Jev statistics to cite?

From this page. Every figure is dated, its method is stated, and the rows behind it are published as JSON at https://madewithjev.com/jev-statistics.json. Quoting it needs no permission, only a link. For a figure that is not here, a cut of the data by category, or a comment on the record, ask Jon Kraayenjon at https://x.com/kraayenjon.

Can this data be reused?

Yes. The figures are free to quote with a link to madewithjev.com. The whole table is also published as JSON at https://madewithjev.com/jev-statistics.json.