Jev Shipped 19 Days Before This Clip, and the 40% Coding-Agent Saving Is Now 28.6% in Its Own Author's Repo

This is a 40-second clip at 1080x1920, posted 2026-10-04 at 14:30:52 UTC by @brochbuilds, whose TikTok channel nickname is "Ben Broch | AI Saas builder", carrying 4,348 views, 150 likes, 28 comments, 18 reposts and 134 saves when the metadata was captured on 2026-10-05 at 05:10:28 UTC, over a track listed only as "original sound" credited to the same account; I read all 20 extracted frames at 2-second intervals and all 157 transcript words. A white title card reading "10 useful ◈ Jev builds" sits over the speaker for the first four seconds, then ten demos stack up under burned-in captions. Build one is a text box: a user types "dinner with priya friday 8pm" and the box resolves into an "Event" card labeled "Dinner" with chips "Friday", "8 PM", "+ Add place", a person chip "P Priya", and footer controls "Esc to clear" and "Add event ⏎". Build two is a dark 3D scene of blue folders and white file cards clustering and re-clustering on a starfield, captioned "Someone built a Google" and "and you can search". Build three is a README screenshot headed "Ask your Postgres tables questions in plain language" with badges "CI passing", "pgxn 0.2.0", "license PostgreSQL" and "website pgjev.com", above a psql session running round(jev_prob(a, 'The story is primarily about programming, software engineering, computer systems, databases, security engineering, developer tools, or building with AI rather than politics, biology, sports, culture, or general news.')::numeric, 3) AS engineering_probability against saved_developer_articles, printing NOTICE: jev: noul + judged 30 rows of saved_developer_articles in 2 requests, 6129 input tokens (≈$0.0003), 2641 ms and a result table topped by "Show HN: Made an open-source Lego AI generator" at 0.970. Build four is an inbox ranked beside sleep data showing "Asleep 4:25 usual 6:58", "HRV 36 usual 25", "Resting 66 usual 69", a free-text box reading "i want to prioritize writing today", and a "Criteria weights" panel listing "Waiting on me up to +0.70", "Sender up to +0.35", "At stake up to +0.35", "Deadline up to +0.50", "Effort up to 0.00 −0.15", "New up to +0.30" and "Shows when the score is ≥ 0.45", with 21 messages under "Worth attention". Build five is a call-center panel marked "BUILT WITH LIVEKIT AGEN..." and "JEV SWITCHBOARD" where turn 1, "Yo, my internet keeps dropping every few minutes. This is so frustrating.", produces "Route → Tech support" under "rule: intent · intent tech_support 1.00 ≥ 0.60" at frustration 2.0 "frustrated", and turn 2, "What do you mean you found the problem, dude? Like, I've been calling you guys so many times. This is the 3rd time this week and I'm just talking to you freaking bots. I'm done. Just get me a real person.", produces "Escalate → human" under "rule: wants_human · wants_human 0.99 ≥ 0.80" at frustration 2.8 "furious". Build six is a signup flow headed "Signup auto enrich / Sign in with Google, the form fills itself" that greets "Welcome, Qure Skincare" and pre-fills "Company size 11-50 / 32 employees on record", "Based in Europe / based in Croatia", "Industry E-commerce / DTC / Jev · 1.00 sure", "Who you sell to Consumers (B2C) / Jev · 1.00 sure" and "You advertise on Meta Google / 428 Meta ads running · 2,000 Google ads running". Build seven is an end-to-end test driving a plant storefront at localhost:4310 called "Fern & Co" with the tagline "Plants that like you back", clicking from the grid into "Calathea Orbifolia" and finishing "Test Files 0 passed / Tests 0 passed / Duration 4.91s" in a repo named e2e-clef-demo. Build eight is a benchmark card labeled "Creator's original benchmark · @dzhng" reading "JEVGREP · jg", "SWE-BENCH · 10-TASK REPEAT · FULL TASK COST", "Codex, prompted to use jg as its researcher.", with "Codex alone BASELINE" struck through at "$7.62", "Codex + jg JEV RESEARCH AGENT" at "$4.52", a red "−40%", and fine print "$7.62 → $4.52 full Codex task cost across 10 tasks, failures included; Jev cost excluded. Solves: 7/10 with jg vs 8/10 baseline. Single repeat on a tuned Python subset, a cost reduction with a quality tradeoff, not a guarantee." Build nine is a two-window screen recording labeled "Voice browser · @moritzkremb" with Wikipedia on the left and a control panel at localhost:8787 on the right showing "voice-browser · Jev", "model jev-1.13.0", "last 300 ms · p50 340 ms", "calls 2", "actions 1", a listening microphone transcript "Jeff underhood okay go to wikipedia.com", and a "JEV DECISION 300 ms · 10 questions" panel resolving to "ACT — open wikipedia.com" with policy gates "is_command 0.94 / 0.5", "intent navigate_url (1) / 0.55", "complete 0.94 / 0.6" and "url_span wikipedia.com / 0.35". Build ten is a quoted X post from "Archive @ArchiveExplorer" under the header "EXCERPT FROM THE CREATOR" reading "jev: scores every review note against DESIGN.md (fix now / later / skip) in milliseconds". The clip closes on two screenshots of the author's own site: a quickstart headed "How to set up Jev and make your first API call in 5 minutes", stamped "Tested September 18, 2026 · By Ben Broch", with stat tiles "642 ms per call, measured", "0.001¢ cost of my first call" and "4 steps, start to finish"; and a resource index reading "27 demos i watched", "71 repos you can run", "35 guides and docs" and "133 resources, updated September 24, 2026".

The product is Jev from TypeSafe AI, and the transcript's "Jeff" is an artifact of speech-to-text

Every spoken instance of the product name in the auto-generated transcript reads "Jeff". That is a mis-hear. The on-screen title card reads "10 useful ◈ Jev builds", the TikTok caption reads "10 useful things people are building with Jev", and the voice-browser panel in frames 2 and 14 shows the model string "jev-1.13.0". The product is Jev, built by TypeSafe AI, a San Francisco company whose site positions it as "System One Models, to be natively used by machines" and "Decisions, not strings". The same mis-hear is visible inside the demo itself: the microphone transcript in the voice browser renders the spoken wake phrase as "Jeff underhood okay go to wikipedia.com", which is the Web Speech API making the identical substitution on live audio.

Jev sits at the model layer rather than being a wrapper over an existing LLM. TypeSafe's launch post, Introducing System One Models & Jev, describes "a new class of frontier models built to make fast, structured decisions that software can use directly", and the public documentation lists three primitives rather than a chat endpoint: Choice ("Choose an option from a list", returning "choice, probabilities, confidence"), Score ("Score the state on a rubric", returning "score, probabilities, confidence") and Noul ("Is this statement true?", returning "noul (0–1)"). That primitive set is visible on screen in the pgjev SQL, which calls jev_prob(...) and logs jev: noul + judged 30 rows.

"Out for two weeks" is wrong in both directions, and neither candidate date lands on fourteen days

The first sentence of the clip claims Jev "has been out for two weeks". The Wikipedia entry for Jev records that it "was released in limited early access on September 15, 2026", alongside a US$40 million seed round led by DCVC at a $200 million valuation. From September 15 to the October 4 post date is 19 days, which is two weeks and five days. TypeSafe's own public introduction post is dated September 28, 2026 and says Jev is "available today in early access", which puts that milestone 6 days before the clip. Fourteen days matches neither anchor. The looser reading of the sentence, that this is still a very young product, holds.

The chronology the clip skips is the part that explains why ten builds existed at all. Vercel moved first on distribution. Vercel posted on X that "Jev was adopted faster than any other model in AI Gateway history. In the first day, @typesafeai reached ~13% of teams, 2x the GPT-5.6 family and 6x Fable 5.1", and by September 18 Jev was reported as the fastest-adopted model in AI Gateway history. The author's own quickstart screenshot in frame 19 is stamped "Tested September 18, 2026" and says "Here's the setup I used to get a real response through Vercel AI Gateway", which places his first call on the same day as that adoption milestone.

The 40% coding-agent saving is a superseded number, and its own author has since published 28.6%

The voiceover says flatly that one build "cuts coding agent costs by 40%". The card on screen is more careful than the voiceover: it is labeled "Creator's original benchmark · @dzhng", and its fine print concedes "Jev cost excluded", "Solves: 7/10 with jg vs 8/10 baseline" and "a cost reduction with a quality tradeoff, not a guarantee". The $7.62 to $4.52 arithmetic checks out at 40.7%, so the number on the card is internally consistent.

It is also out of date. The current dzhng/jevgrep repository reports a baseline of $7.62 against $5.44 for jevgrep, "a measured 28.6% reduction, rounded to ~30%", on "ten tuned Python SWE-bench tasks", with 8 of 10 tasks solved by both approaches rather than the 7-versus-8 split shown in the clip. Counting Jev's own API spend, the 0.4.3 total-cost rerun measured 25.8% lower total cost. The author said as much publicly when he shipped the revision: "jevgrep v0.4 released! ... it is now intelligence parity with codex/claude's own subagents but with 30% less cost. Yes it's less than the original's 40%, but I traded more intelligence for a bit more cost". A later 0.5 release post reports Jev's own cost down 59% while "the per task cost stayed the same". The honest version of the voiceover claim is that jevgrep cut Codex task cost by roughly 29% at intelligence parity, or by 26% once you pay for Jev.

Two of the ten builds are public repositories whose numbers match what the frames show

The voice browser in frames 2 and 14 is moritzkremb/jev-voice-browser, described in its own repository as "Control a real browser by voice. Jev (TypeSafe System One) decides intent + target in ~300 ms per spoken word; Playwright acts". The repo documents average Jev latency of "≈ 330 ms (p50 ≈ 300 ms)", which matches the on-screen readout "last 300 ms · p50 340 ms" within a rounding step, and confirms the model string jev-1.13.0 and a cost of roughly $0.0002 per call. The repo also lists limits the clip does not mention: the Web Speech API works only in Chrome and Edge and sends audio to Google, the element snapshot is capped at 100 items, only one action runs per utterance, iframe contents are unreachable, and sites with heavy bot protection fail to render.

The Postgres build is pgjev, a PostgreSQL extension by realZachi released under the PostgreSQL license, which exposes jev(), jev_prob(), jev_choice() and jev_score() as SQL functions and batches up to 40 rows per API request. Its published figures are about 3.5 seconds and roughly $0.012 for a first pass over 2,000 rows, dropping to about 50 milliseconds on cached runs. The cost line burned into frame 5, "6129 input tokens (≈$0.0003)", is consistent with TypeSafe's posted input price of "$0.042 / MTok ($42 per billion tokens)", which works out to $0.000257 for 6,129 tokens.

The call-center build is labeled on screen as "BUILT WITH LIVEKIT AGEN..." and matches a published Jev Switchboard demo in which Jev judges intent, frustration, whether the caller wants a human and whether the caller is trying to manipulate the agent, with plain code rules routing or escalating. The remaining builds in the clip, including the self-organizing file visualization, the sleep-weighted inbox, the signup auto-enrich flow and the Fern & Co end-to-end test, carry no attribution label in any frame. I could not match them to a specific public repository, so I am not naming authors for them.

TypeSafe's headline speed and cost multiples come from workflows TypeSafe wrote itself

The clip does not quote TypeSafe's marketing numbers, but every build in it inherits them, so they are worth pinning down. TypeSafe's home page advertises "193.6x Faster, 444.6x Cheaper", "Completed in 0.114s vs. 8.566s for LLMs", "$0.000081 vs. $0.013880" per comparable workflow, "$42 Per Billion input tokens" and "238x Lower input price than Claude Fable 5.1". Output tokens are listed as "FREE (too cheap to meter)". The launch post states "End-to-end response time is 70ms-500ms" and a general claim of "40x-200x faster".

Wikipedia records a caveat TypeSafe itself supplied: the benchmark workflows "were created by members of its model-capabilities team", and the company describes "the reported gains as likely to sit at the high end of real-world results". The clip's own evidence supports that caution. The author's quickstart card reports "642 ms per call, measured", which sits above the top of TypeSafe's published 70 to 500 millisecond band. The voice browser's repo separately notes an initial TLS handshake of around 700 ms before the fast path kicks in.

The resource offer in the call to action does not count what the voiceover says it counts

The voiceover ends with "a list of over 100 Jev use cases"; the TikTok caption says "130+ use cases". The page shown in frame 20 says something different. Its three tiles read "27 demos i watched", "71 repos you can run" and "35 guides and docs", summing to exactly 133, and the page itself labels the total "133 resources, updated September 24, 2026". Those are 133 links across three categories, not 133 distinct use cases, and 71 of them are repositories rather than documented applications. The search placeholder on the same page reads "Search 133 resources: lead scoring, invoices, n8n...".

The index is also nine days stale relative to the post. It is stamped September 24, 2026 while the video went up October 4, 2026. Two of the three featured guide cards carry their own measured claims, "586 pages for 21¢" for SEO internal linking and "6 buyers in 30 posts, tested" for prospect finding, neither of which I could trace to a public source outside the author's own site.

Key Takeaways

• Correction: the clip's "out for two weeks" does not match either anchor date. Jev entered limited early access on September 15, 2026, which is 19 days before this October 4 post, and TypeSafe's public introduction post is dated September 28, 2026, which is 6 days before it.
• Correction: the transcript's "Jeff" is a speech-to-text error for Jev, the model from TypeSafe AI. The same substitution happens inside the voice-browser demo on screen, where the live transcript reads "Jeff underhood okay go to wikipedia.com".
• Partial correction: "cuts coding agent costs by 40%" is a superseded figure. jevgrep's own repository now reports $7.62 to $5.44, a 28.6% reduction at 8 of 10 tasks solved for both arms, and 25.8% once Jev's own API cost is counted. The author posted the revision himself and called the original 40% a trade of intelligence for cost.
• Partial correction: the 40% card on screen is honestly footnoted even though the voiceover is not. The fine print states "Jev cost excluded" and "Solves: 7/10 with jg vs 8/10 baseline", meaning the original 40% came with one fewer task solved.
• Verified: the voice browser is moritzkremb/jev-voice-browser running model jev-1.13.0, and the repo's published "≈ 330 ms (p50 ≈ 300 ms)" matches the on-screen "last 300 ms · p50 340 ms".
• Verified: the Postgres build is pgjev, PostgreSQL-licensed, exposing jev(), jev_prob(), jev_choice() and jev_score(). The frame's "6129 input tokens (≈$0.0003)" reconciles with TypeSafe's posted $42 per billion input tokens, which gives $0.000257.
• Verified: TypeSafe publishes input at "$0.042 / MTok" with output "FREE (too cheap to meter)", and an end-to-end band of "70ms-500ms".
• Verified: Vercel reported Jev as the fastest-adopted model in AI Gateway history, reaching roughly 13% of teams on day one, twice the GPT-5.6 family and six times Fable 5.1.
• Unstated cost: Jev is a paid, proprietary, hosted API behind an early-access waitlist. Nothing in the clip runs locally, and every build shown adds a per-call bill. pgjev quotes about $0.012 for a first pass over 2,000 rows; the voice browser quotes about $0.0002 per call.
• Unstated caveat: TypeSafe's own 193.6x and 444.6x headline figures come from workflows written by its model-capabilities team, and the company describes them as sitting at the high end of real-world results. The author's own measured 642 ms per call is above TypeSafe's published 500 ms ceiling.
• Unstated caveat: the "over 100 use cases" offer resolves to a page counting "133 resources" as 27 demos plus 71 repos plus 35 guides, last updated September 24, 2026, nine days before the video posted.
• Unverified: the self-organizing file visualization, the sleep-weighted inbox, the signup auto-enrich flow and the Fern & Co end-to-end test carry no attribution in any frame, and I could not match them to specific public repositories.

Resources

• TypeSafe AI home page. Vendor site for Jev, source for the "$42 Per Billion input tokens", "238x Lower input price than Claude Fable 5.1" and "193.6x Faster, 444.6x Cheaper" claims.
• Introducing System One Models & Jev. TypeSafe's launch post, dated September 28, 2026, with the "70ms-500ms" latency band and the free-output pricing.
• TypeSafe AI documentation. The three API primitives, Choice, Score and Noul, with their return shapes.
• Jev (AI model) on Wikipedia. September 15, 2026 early-access date, the $40M DCVC seed at a $200M valuation, and TypeSafe's own caveat that benchmark workflows came from its model-capabilities team.
• dzhng/jevgrep. The current benchmark, $7.62 to $5.44 at 28.6%, 8 of 10 tasks solved both ways, 25.8% once Jev's cost is included.
• dzhng on jevgrep v0.4. The author stating the new figure is 30% and explaining why it is below the original 40%.
• dzhng on jevgrep 0.5. Jev cost down 59% with per-task cost unchanged.
• moritzkremb/jev-voice-browser. The voice browser in frames 2 and 14, with model jev-1.13.0, p50 latency near 300 ms, roughly $0.0002 per call, and its documented limits.
• pgjev. The PostgreSQL extension in frame 5, with its four SQL functions, 40-row batching, and 2,000-row cost and timing figures.
• Vercel on Jev adoption. The day-one AI Gateway adoption numbers quoted above.
• TypeSafe AI's Decision Model Jev Becomes Vercel's Fastest Adopted Launch. Secondary reporting on the September 18 adoption milestone.