JEV

Independent · Measured · No promotion

TypeSafe Jev,
observed.

Public usage cases grouped by the decision they make, plus latency and cost measurements we ran ourselves on an identical task set. We do not verify what authors claim, and we do not make a page for every post.

Browse 336 casesSee our measurements

Case index

Public builds, grouped by the kind of decision they make. Nothing republished.

336

cases indexed

Benchmarks

4 decision tasks, 3 models, every raw response published.

396

measurement calls

Decision types

8 hubs
92 cases

agent loop

Tool, route and guardrail choices mid-loop.

87 cases

document classification

One call per page, row or record.

76 cases

browser agent

Element-level choices inside a live page.

72 cases

game decisions

Enumerated legal moves under a frame budget.

48 cases

inbox triage

A decision per message, at volume.

30 cases

trading decisions

Rules in code, judgement in the model.

25 cases

supply chain gate

Allow or refuse, before the code runs.

10 cases

text detection

A probability instead of a classifier.

One number we stand behind

Full method →

Cost per 1,000 decisions · warm n=124 per model

lower is better
Jev (TypeSafe)$0.0200acc 96.0%
DeepSeek V4.1 Flash$0.0351acc 94.4%
DeepSeek V4 Pro$0.1497acc 96.0%

Jev is the cheapest here, but only by about 1.8× against the cheap baseline — not the 40–400× you see quoted. Against the frontier baseline the gap is about 7×. Accuracy is shown alongside because on these four tasks all three sit at ceiling and accuracy separates nothing.

Warm p50 latency

312.7 ms

against 263.4 ms and 258.8 ms for the two baselines.

Round-trip floor

281–944 ms

A minimal request already costs this much from where we measured. At that distance a single call is dominated by the network, not the model.

Featured builds

Six of them, with the numbers the authors published. These are shown through X's own embed, so the post and its media are served by X rather than copied here.

6 of 336
  • browser agentofficial embed

    A flight search end-to-end, published at 1× speed with the cost attached.

    Open on x.com ↗
  • document classificationofficial embed

    Filenames stripped, so only content could decide the label.

    Open on x.com ↗
  • game decisionsofficial embed

    The method is in code; the model only picks the next step.

    Open on x.com ↗
  • agent loopofficial embed

    The two-stage design: classify cheaply, escalate only the uncertain tail.

    Open on x.com ↗
  • trading decisionsofficial embed

    Rules in code, judgement in the model, one decision per block.

    Open on x.com ↗
  • browser agentofficial embed

    A real-time loop costing fractions of a cent per item.

    Open on x.com ↗

Embeds load X's widget script. If it is blocked, the post text and a link to the original still render.

Recently indexed

All 336 →
  • 2026-09-19@kushwho11146Hey Everyone, Just Built : Audit your git diff against YAML coding-standards packs using TypeSafe's Jev model, from a CLI or your AI agent's command/skill. 1 step install: npx @kushwho/jev-codes…
  • 2026-09-19@jiayaoctrl-f has never understood a single query. It matches strings. This one scores meaning: every sentence on the page gets a probability with @typesafeai's Jev, and walks you through the hits.…
  • 2026-09-19@acharyaagamyaI made a Magic Jev Ball for code reviews 🎱 Click it on any GitHub PR and ask: "should I approve this?" It checks CI, diff size, and reviews, then lets @typesafeai Jev decide your fate in ~200 ms No…
  • 2026-09-18@maubaronjev is insane 🤯 here is jev playing smash bros against itself he is controlling all 4 different characters. and literally deciding whats the best move to play against itself all within a fraction of…
  • 2026-09-18@wuyang_zhoubtw this is what you get if you ask Jev to play Minecraft with GPT-6 Astra. I still can’t believe it
  • 2026-09-18@wiley_laddTypeSafe AI is challenging the industry’s reliance on large language models for every stage of the agentic stack with the launch of Jev, a specialized “System One Model” designed exclusively for…

What this site is

Two parts, and they support each other. The case index collects public builds and groups them by decision type, so you can find the one that looks like your problem rather than the one with the most views. The benchmark page runs Jev and two chat models on an identical task set and publishes every raw response, so you can recompute the numbers instead of trusting them.

The grouping is the useful part of the index. “Someone used it to sort files” and “someone used it to drive a browser” are not the same kind of fact; what transfers between projects is the shape of the decision, not the domain.

What this site is not

It is not affiliated with TypeSafe, and it is not a mirror of anyone’s posts. Every case links to its original post, keeps only a short excerpt, and rehosts no media and no threads. Where a claim comes from an author we say so and quote their figure as reported rather than verified.

It is also not a tutorial library. The concepts of the model — what the three primitives are, how confidence thresholds work, when not to use it — are already documented well elsewhere, and we link out instead of writing them again.

Start here

If you want to know whether it is worth it, read the benchmark page first and run the scripts from your own region. If you already know what you want to build, start with the case index or the hub closest to your problem: browsers, games and realtime control, or document classification.