Free output tokens turn classification into a product, but TypeSafe's own docs rule out two of these seven businesses

Jev's pricing removes the cost ceiling on per-item classification, which is the real argument underneath all seven ideas in this video. The clip is a 1:46 talking-head piece from Kevin at @vibewithkevin, shot in a home office in front of a lit bookshelf with a copy of Dune and a DJI lav mic clipped to his shirt. A yellow-highlighted banner reads "STEAL THESE 7 JEV BUSINESS IDEAS" and word-by-word burned-in captions run across the lower third. Seven numbered cards slide in over the footage, each with the same three fields: BUILD, BUYER, and a highlighted PRICE line. One extra card, labeled RECEIPT, quotes a cost comparison from an X post. The claim is that TypeSafe AI's Jev model is cheap enough to classify every record in a dataset instead of a sample, and that seven small businesses fall out of that.

What Jev is, checked against the vendor

The video's one-line definition holds up. TypeSafe's launch post, Introducing System One Models & Jev, describes a model that takes program state plus typed questions and returns typed decisions in a single parallel pass. It answers with a Choice, a Score, or a boolean probability, each carrying a probability distribution and a confidence value. MarkTechPost's write-up puts it plainly: "It does not generate text."

The dates and numbers in the video are accurate. TypeSafe launched Jev on September 15, 2026. Input is priced at $0.042 per million tokens and output is free, which TypeSafe labels "too cheap to meter." The company quotes end-to-end latency of 70ms to 500ms and headline multipliers of "193.6x faster, 444.6x cheaper" against LLM baselines on its own workflow evaluations. Kevin's "4.2 cents per million tokens and output is literally free" is an exact match to the published price. His "milliseconds" claim for the agent guardrail idea sits inside the published latency band.

Two details the video leaves out matter for anyone acting on it. Jev's context window is 32,000 tokens, per the Vercel AI Gateway model page, which caps how much state you can hand it per decision. And TypeSafe says on the launch post that the price may be subsidized and is expected to change.

The $1.50 receipt

The RECEIPT card quotes @marcoporracin, dated Sep 17, with 69.3K views, and reads: "4. Costs and time (35k items). Gemini 3.8 Flash: ~US$ 23 · ~4 pairs/s · hours (ceiling ~300 RPM). Jev 1.13: ~US$ 1.50 · ~150 pairs/s · 4 min." I could not find that post or that account through search, so I am treating the specific figures as unverified. The surrounding facts do check out. Jev 1.13 is a real version, released September 18, 2026, priced at $0.042 per million input tokens. Gemini 3.8 Flash is a real model, and Google DeepMind's model card lists $0.75 per million input tokens with caching and $3.75 per million output.

The receipt's internal arithmetic is consistent. At $0.042 per million, $1.50 buys roughly 35.7 million input tokens, which works out to about 1,020 tokens per pair. That is a reasonable size for two product listings. The same token volume through Gemini 3.8 Flash at $0.75 per million would run about $27 on input alone, so a $23 measured figure implies a slightly leaner prompt on the Gemini side. The cost ratio in the receipt is 15.3x; the published per-token input ratio is 17.9x. Those agree closely enough that the number is plausible. The timing also reconciles: 35,000 items at 4 pairs per second is roughly 2.4 hours, and at 150 pairs per second it is 3.9 minutes.

An independent test corroborates the order of magnitude. A developer blog post, The Thirty-Cent Judge, ran 9,081 low-confidence product pairs through Jev for $0.32 in 13 minutes 22 seconds at 6 concurrency, sending each pair as two named objects plus the first 1,500 characters of a scraped description. That is about $0.000035 per pair, against the video's implied $0.000043 per pair. The throughput gap is a concurrency artifact, not a model difference.

So the cost claim survives. The 150 pairs per second figure depends entirely on how many parallel requests the tester ran, and the video presents it without that context.

Where the seven ideas meet the model's documented limits

TypeSafe publishes a page called Jev 1.13 jaggedness listing the failure modes it knows about. Jev "does not count reliably," is "not a calculator," "reads dates as text, not as ordered quantities," and loses accuracy "as the state grows with content unrelated to the decision." It also warns that "content written to adversarially steer the model" can move the answer.

Three of the seven ideas line up cleanly with what the model does well. Catalog cleanup (idea 2) is semantic comparison of two records with a human review lane for low-confidence pairs, which is exactly the shape the paddo.dev post implemented in about 150 lines. Email triage (idea 7), listed on the card as "sort and rank, write nothing," is routing plus scoring on bounded text. Agent guardrail (idea 5), "approve, block or escalate every command," is a three-way Choice on a short command string, and the latency band supports it.

Two run into trouble. The AI text detector (idea 1) inherits a problem that has nothing to do with Jev. A 2023 evaluation, Testing of Detection Tools for AI-Generated Text, found detection tools are "neither accurate nor reliable," and the Wikipedia summary of the field collects the false-positive research, including findings that detectors flagged a majority of genuine TOEFL essays by non-native English speakers as machine-written. Selling that to "teachers, editors, hiring" is selling a scoring product whose errors land on individual people. Lead scanning (idea 4) is a filtering job over social posts, which is a fit on paper, but the posts are exactly the kind of adversarial, marketing-shaped text TypeSafe flags, and the volume of irrelevant surrounding content works against the model's documented sensitivity to noisy state.

Agent memory (idea 6) is the most interesting of the seven and the least proven. Picking which stored memories are relevant right now is a ranking problem over a growing store, and Jev's 32,000-token context window means you are paging through candidates rather than scoring a full memory bank in one call.

The accuracy question the pricing does not answer

Cheap classification is only useful if the classifications are right. JevBench, which states it runs with "no sponsorship, no free credits, no pre-publication review," reports Jev at 80.3% on Banking77 intent classification where supervised encoders reach 93.3%, 62.6% on phishing detection against Claude Haiku 4.5 at 81.3%, and 87.0% on CLINC150 routing against GPT-5.6 Terra at 91.5%. On email classification, Jev variants landed between 93.6% and 98.6% while a TF-IDF ensemble hit 99.3%.

That is the missing half of the video's pitch. For the search reranker (idea 3), where a wrong call quietly degrades results, and for the lead scanner, where a false positive wastes a salesperson's morning, the accuracy trade is the whole decision. MarkTechPost makes the same point about TypeSafe's zero-hallucination language: "'Zero hallucinations' means schema matching is guaranteed. Answers can still be wrong." The guarantee covers the shape of the response, not its correctness.

The closing advice expired one day after upload

Kevin ends by telling viewers stuck on the TypeSafe waitlist that Jev is live on Vercel's AI Gateway. That was correct when he posted. Vercel's changelog entry dated September 16, 2026 confirms the model is callable as typesafe-ai/jev through AI SDK 7's experimental_evaluate function. The video went up September 20. On September 21, TypeSafe posted on X that "Jev is now available to everyone. No waitlist." The workaround is now unnecessary, which is a fair thing to happen to a video about a six-day-old model.

Key Takeaways

• Jev's published price is $0.042 per million input tokens with free output, launched September 15, 2026, with a 32,000-token context window and 70ms to 500ms latency. The video quotes all of this correctly.
• The $1.50 versus $23 receipt is internally consistent with published per-token pricing and is corroborated in order of magnitude by an independent product-matching test that ran 9,081 pairs for $0.32.
• The 150 pairs per second throughput figure depends on the tester's concurrency setting, which the video does not mention.
• TypeSafe's own jaggedness documentation rules Jev out for counting, arithmetic, date ordering, and noisy state, which weakens the lead scanner idea and does nothing for the AI text detector's underlying accuracy problem.
• Independent benchmarks put Jev below supervised classifiers and frontier LLMs on several classification tasks, so the cost advantage comes with an accuracy cost the video does not raise.
• Whisper's transcript is accurate except for two name errors: it renders "Jev" as "JEV" throughout and as "Jeff" in the final line. The burned-in captions and the on-screen cards spell it "Jev" and "TYPESAFE" correctly, and the video description confirms the spelling.
• Unverified: the @marcoporracin post containing the 35,000-pair receipt. I could not locate the account or the post. The card's figures are arithmetically plausible and match published pricing ratios, but I have no primary source for them.
• Unverified: the claim that Gemini 3.8 Flash has a ~300 RPM ceiling. Google's model card does not publish rate limits.

Resources

• Introducing System One Models & Jev -- TypeSafe's launch post, the primary source for the September 15 date, the $0.042 input price, free output, and the 40x-200x speed claims.
• Jev 1.13 jaggedness -- TypeSafe's own list of what the model is bad at, including counting, math, date comparison, and noisy state. The most useful page for deciding whether any of these seven ideas is buildable.
• TypeSafe AI's Jev now available on AI Gateway -- Vercel changelog from September 16, 2026 confirming the typesafe-ai/jev model id and the experimental_evaluate call pattern the video points viewers toward.
• Jev API, Pricing & Playground -- Gateway model page with the 32,000-token context window and the $0.042 per million input price.
• The Thirty-Cent Judge -- Independent product-matching run: 9,081 pairs for $0.32 in 13 minutes. The closest real corroboration of the video's catalog cleanup economics.
• JevBench benchmarks -- Third-party accuracy comparisons against supervised encoders and frontier LLMs, with a stated no-sponsorship policy.
• TypeSafe AI Releases Jev -- Independent coverage that separates the type-safety guarantee from answer correctness and flags the vendor-run nature of the launch benchmarks.
• Gemini 3.8 Flash model card -- Google DeepMind's published pricing, used to check the receipt card's $23 figure.
• Testing of Detection Tools for AI-Generated Text -- Peer-reviewed evaluation finding AI text detectors are "neither accurate nor reliable," relevant to idea 1.
• TypeSafe AI on X, September 21, 2026 -- The post announcing the waitlist removal that made the video's closing tip obsolete.