<- all tokdocs
2026.09.20@itsborjao

Jev did not read 586 pages in 45.1 seconds: the dashboard labels itself a "simulated run" at "2x real" speed, and TypeSafe's own docs say Jev does not browse the web

AI
Watch on TikTok

View on TikTok ->

A 27.84-second vertical clip (video.mp4, 1,774,302 bytes, QuickTime/MP4 container reported by ffprobe as mov,mp4,m4a,3gp,3g2,mj2, overall bitrate 509,875 bps; video stream HEVC Main profile level 120, tag hvc1, 576x1024, yuv420p, 30/1 fps, 835 frames, 438,605 bps; audio stream AAC HE-AAC, 44.1 kHz stereo, 64,270 bps, 602 frames, 27.839002 s) posted by uploader itsborjao (uploader_id 7671656094218339341, channel nickname "borja") at epoch timestamp 1789917690, which is 2026-09-20T15:21:30Z, credited to "original sound" by borja, with an 858-character description. At the capture time of metadata.json (Oct 19 2026 06:17:31 local) the counts read 2,240 views, 95 likes, 1 comment, 24 reposts and 87 saves. I read frames 001, 003, 005, 007, 011, 012, 013 and 014 from tokdoc_frames/ plus targeted high-magnification crops rendered straight from the video at t = 0.5, 6, 14, 20, 21.9, 22.5 and 23.1 s, against a 6-segment, 84-word Whisper transcript. The top half of every frame through t ≈ 23 s is a screen recording of a web app titled "Internal Link Dealer", subtitled "by Distribb", with a header line reading "Edition of September 1X, 2026 · 586 pages crawled off distribb.io · 15 destinations · 8790 page pairs · simulated run" and a control strip at top right reading "Pages [586] Speed [2x real ▾] Sound off [Stop]". Two columns race: a pink "Jev" column labelled "READING AND MAPPING" and a cream "Claude Opus 5" column labelled "anthropic · via Bedrock/Anthropic · READING AND MAPPING". The Jev counters run 24 read / 27 mapped / 1,746 judgments / $0.0056 at elapsed 1.5 s, 165 / 161 / 16,303 / $0.0584 at elapsed 12.7 s, 378 / 363 / 36,072 / $0.1348 at elapsed 28.7 s, 533 / 533 / 51,723 / $0.1933 at elapsed 40.7 s, and freeze at 586 read, 573 mapped, 56,844 judgments, 1261 judgments per second, elapsed 45.1, cost $0.2111; the Opus column freezes at 21 read, 18 mapped, 1,809 judgments, 40 judgments per second, elapsed 45.1, cost $1.43. The lower half is a talking head in a cap and an orange-and-navy Red Bull KTM jacket, with burned-in word-by-word captions ("JEV is WILD for SEO Audits 🤯", "co-inventor", "and", "cost", "I'll", "plugin"). At frames 013 and 014 (t ≈ 24-27 s) the shot switches to a phone camera pointed at a monitor showing a ChatGPT settings panel (URL fragment ...3c-f9a8-83e9-a2f8-e6d048389f19#settings) with a sidebar reading General, Notifications, Personalization, Plugins, Voice, Billing, Usage, Analytics, Data controls, Cloud browser, Storage, Safety, Security and login, then a "Plugins" pane headed "Work with ChatGPT across your favorite..." listing "Able SEO by WixSEO", "SEO Programatico", "Semantic SEO Evidence", "S... with Distribb", "...Pilot Lite" and "Content Engineering". No terminal, no API response, no log file and no crawler output appears in any frame I read: every number in the video comes from this one self-described simulated dashboard. Spoken, the argument is that Jev is a new fast model from the co-inventor of ChatGPT that read 586 pages and rebuilt an internal link map in 45 seconds for 21 cents while Claude Opus 5 managed 21 pages for $1.43, and that the creator will ship it as a ChatGPT plugin. The transcript's last segment ends at 27.76 s against an audio stream of 27.839002 s, so nothing is timestamped past the end of the audio and no segment is hallucinated.

"It read through all 586 pages on my website" is wrong about what Jev does: TypeSafe's docs say Jev does not browse the web

At 6.08 s the video says "It read through all 586 pages on my website and revealed the internal link map in 45 seconds and only cost 21 cents." Whisper renders the model name as "Jeff"; the burned-in caption at frame 001 reads "JEV is WILD for SEO Audits" and the dashboard column header reads "Jev", so the subject is Jev, not Jeff.

TypeSafe's own model reference states the limit plainly. docs.typesafe.ai/models says Jev "does not generate text or browse the web" and that it "reads natural-language text, sent as a string, a JSON object, or an array of text values." Jev takes state plus questions and returns typed decisions with probabilities. It has no fetcher, no crawler and no HTTP client of its own.

The video's own dashboard agrees. The header line says "586 pages crawled off distribb.io", past tense, as a property of the edition being replayed, separate from the race underneath it. The crawl was done by Distribb's harness beforehand. What the 45.1-second clock measures is the decision pass over an already-fetched corpus. Saying the model "read through all 586 pages on my website ... in 45 seconds" merges a crawl that happened earlier with a classification pass that happened later, and attributes both to Jev.

This matters for the plausibility question the headline numbers invite. 586 pages fetched in 45.1 seconds is 12.99 requests per second sustained against one origin. Nothing in the video shows that happening, and nothing needs to have happened, because the pages were already on disk.

The dashboard labels itself "simulated run" and the clock is explicitly running at "2x real"

Two pieces of on-screen text undercut the framing of a live race. The header's final token, resolved by temporally averaging 128 frames of the static region and upscaling, reads "simulated run". The control strip at top right carries a speed selector whose current value reads "2x real", next to a "Stop" button and a "Pages" field showing 586.

The frame timings confirm the 2x figure independently of my reading of the label. The Jev elapsed counter reads 1.5 at video t = 0.5 s and 45.1 at video t = 22.5 s: 43.6 seconds of dashboard clock across 22.0 seconds of video, a ratio of 1.98. That is why a run advertised as 45.1 seconds fits inside a 27.84-second clip.

A simulated replay at 2x is not the same artifact as a captured live benchmark, and the video never says which it is. The spoken line "in 45 seconds" and the title's "45.1 seconds" both present the dashboard clock as wall time for a real run. Readers should treat the whole comparison as a replay of a prior result, rendered by the vendor of one of the two products in the race, with no raw output shown.

Jev is real, Diogo Almeida is real, and "co-inventor of ChatGPT" is press framing rather than a verifiable title

At 2.16 s the video says Jev is "a new super fast AI model developed by the co-inventor of ChatGPT." The underlying facts hold up, with one qualification.

Jev is TypeSafe AI's first public model. Artificial Intelligence News, September 16, 2026 reports TypeSafe left stealth on September 16, 2026, gives 70 to 500 ms end-to-end latency, $0.042 per million input tokens with no output charge, and a "hardware-aware parallel sampler with type-safe structured outputs (up to 255 discrete options)", and quotes the founder as Diogo Almeida: "Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." TypeSafe's own site describes Jev as "TypeSafe's first public System One Model, optimized for automation."

The name spelling is Diogo Almeida, not Diego. Secondary coverage splits on this, so I checked it against a primary source: the arXiv listing for Training language models to follow instructions with human feedback, arXiv:2203.02155, submitted March 4, 2022 lists Diogo Almeida as the fourth author, after Long Ouyang, Jeff Wu and Xu Jiang. That paper is InstructGPT, the RLHF work ChatGPT was built on. "Co-author of the InstructGPT paper" is checkable; "co-inventor of ChatGPT" is a press shorthand for the same person and is not a title any primary source assigns.

Jev is also not open. TypeSafe ships it as a hosted API, and has not released weights or training data. The repositories that turn up under names like openjev are third-party reimplementations of the interface on open base models, not the model: razorback16/openjev is Apache-2.0, and TheoLeeCJ/SemIf-OpenJev, sabeel111/OpenSourceJev, Zefan-Cai/Open-Jev and GPT-AGI/OpenJev are MIT. None of them is TypeSafe code. If you want Jev itself, there is one supplier and one price.

The version matters too. docs.typesafe.ai/models lists Jev 1.14 (jev-1.14.0) as released October 14, 2026, with jev-latest still resolving to Jev 1.13.0. This video was posted 2026-09-20, so the run shown can only have used 1.13.

The cost numbers: Jev's $0.2111 reconciles against published pricing, the "$43 full Opus pass" does not

The two cost figures on screen are checkable against both vendors' list prices.

Jev's side closes cleanly. TypeSafe charges $0.042 per million input tokens with output free. The displayed $0.2111 buys 5.026 million input tokens. Divided across 56,844 judgments that is 88.4 input tokens per judgment, which is a plausible budget for a short page-pair question. The judgments-per-second column is also internally exact: 56,844 / 45.1 = 1,260.4, and the column reads 1,261. On the Opus side, 1,809 / 45.1 = 40.1, and the column reads 40. Whatever else the dashboard is doing, those two columns are consistent with each other.

The Opus figure is harder to pin. Anthropic's Claude Opus 5 model page lists $5 per million input tokens and $25 per million output tokens, released July 24, 2026, now marked "Legacy", with adaptive thinking and a default effort of high. If the displayed $1.43 were spent entirely on input, it would buy 286,000 tokens, or 158 tokens per judgment across 1,809 judgments. That leaves almost no room for output: even 200 output tokens per judgment would cost 1,809 x 200 / 1,000,000 x $25 = $9.05, more than six times the figure shown. So the Opus column implies a near-zero-output configuration of a model whose documented default is adaptive thinking at high effort. I could not resolve the fifth column's label at 576p, so I am not treating it as a contradiction, only noting that the Opus spend shown is low relative to the model's documented default behaviour.

The 190x claim does check out. The description says "per page that is ~190x cheaper". Jev: $0.2111 / 586 = $0.00036024 per page. Opus: $1.43 / 21 = $0.06810 per page. The ratio is 189.0. That is the one headline multiplier in this video that survives arithmetic.

The projection attached to it does not. The description says "the full Opus pass would have run $43." At the Opus rate the dashboard itself displays, $0.06810 x 586 = $39.90. The $43 figure is 7.8% above what the screen supports, and I could not derive it from any combination of the displayed counters.

The description's own tallies do not match the dashboard it is describing

The description makes three specific claims about the result. All three disagree with the frozen counters in the same frames.

It says "584 links placed". The Jev MAPPED counter freezes at 573, confirmed at t = 22.5 s and again at t = 23.1 s.

It says "139 pages it refused to link because nothing honestly fit". The screen shows 586 read and 573 mapped, leaving 13 unmapped. 139 and 13 are the same digits in a different order.

It says the job is "8,790 yes/no calls", and the dashboard header agrees, reading "8790 page pairs", which is exactly 586 pages x the "15 destinations" also named in that header. But the Jev JUDGMENTS counter freezes at 56,844, 6.47 times the stated 8,790. Either the judgment counter is measuring something other than page pairs, or the stated workload is understated by a factor of six. The video does not say which, and the cost arithmetic in the previous section depends on the answer: at 8,790 calls instead of 56,844, the $0.2111 implies 572 input tokens per call rather than 88.

None of these is fatal to the demo. Together they show that the description was written against a different set of numbers than the ones rendered in the clip.

Opus 5 is the wrong Claude model for this job, and Anthropic's own docs say so

The comparison is Jev against Claude Opus 5. Anthropic's models overview describes the current lineup by intended workload. Claude Opus 5.5 is "For long-running agentic coding and knowledge work". Claude Haiku 5.5 is "For high-volume, latency-sensitive tasks such as classification, extraction, and routing" and is priced "From $0.10 / input MTok, From $0.50 / output MTok".

The description's own framing of the task is "8,790 yes/no calls ... that is a classification problem". That is the exact workload Anthropic routes to Haiku, at one fiftieth of Opus 5's $5 per million input tokens. Racing a decision model against Opus 5 on a yes/no classification job compares against the most expensive model in the family rather than the one documented for the task. The Batch API would cut the Opus side by another 50% (Opus 5 batch: $2.50 input, $12.50 output), and a batched classification job has no latency requirement that would rule it out.

The description also discloses a methodology detail the spoken version omits: "the run stops the moment Jev finishes so Opus stops burning tokens." So Opus's 21 pages is a cutoff, not a ceiling. The spoken line at 14.76 s, "Cloud Opus 5 on the same amount of time could only handle 21 pages and cost $1.43", is fair about the time box, and Whisper's "Cloud" is its usual mangling of "Claude", confirmed by the on-screen column header "Claude Opus 5".

Crawl-delay has no standing in the robots.txt standard, so the 13 requests per second question has no spec answer

If anyone does try to fetch 586 pages in 45 seconds, the relevant norm is thinner than SEO folklore suggests.

RFC 9309, Robots Exclusion Protocol, September 2022 defines no Crawl-delay directive at all. It sets a parsing limit ("The parsing limit MUST be at least 500 kibibytes"), a caching rule ("Crawlers SHOULD NOT use the cached version for more than 24 hours, unless the robots.txt file is unreachable"), and a failure rule ("If the robots.txt file is unreachable due to server or network errors ... the crawler MUST assume complete disallow"). It sets no request rate. Google's own robots.txt specification supports only user-agent, allow, disallow and sitemap, and lists crawl-delay among the fields it ignores, managing crawl rate algorithmically instead.

The site in question does not set one either. I fetched https://distribb.io/robots.txt directly: it is 348 bytes, allows all paths for * except /directory/, /login, /signup, /integrations, /seo-audit, /settings, /api/, /dashboard, /content-calendar, /sequences, /campaigns and /inbox, blocks SemrushBot entirely, and declares Sitemap: https://distribb.io/sitemap.xml. There is no Crawl-delay line. That sitemap, fetched the same day, is 93,799 bytes and contains 696 <loc> entries, which is the right order of magnitude for the 586 pages the video claims a month earlier and is consistent with a site that added pages in between.

So there is no robots.txt rule that 13 requests per second would violate on this site, including for the operator's own crawler on his own domain. The constraint on the Jev side is a different one: TypeSafe's docs say rate limits are per organization, "measured in tokens per second and requests per second", with a 429 on exceed, and do not publish absolute numbers. 56,844 judgments in 45.1 seconds would be 1,260 requests per second if each judgment were its own call, which no published tier supports. The docs show why it need not be: "There is no separate limit on the number of questions in a request. Any number of questions is allowed as long as the request fits both limits", with a 64k per-request context and a 32k per-question context. Batching many questions into each request is the documented path to that throughput.

Key Takeaways

  • Correction: The video says at 6.08 s that Jev "read through all 586 pages on my website ... in 45 seconds". TypeSafe's model reference says Jev "does not generate text or browse the web". The crawl was done beforehand by Distribb's harness, as the dashboard header's own phrase "586 pages crawled off distribb.io" indicates; the 45.1 seconds covers the decision pass only.
  • Correction: The dashboard header's final token reads "simulated run", and the control strip's speed selector reads "2x real". The elapsed counter advances 43.6 dashboard-seconds across 22.0 video-seconds, a measured ratio of 1.98, confirming the 2x label. This is a replay, not footage of a live race.
  • Correction: The description says "584 links placed" and "139 pages it refused to link". The frozen counters read 573 mapped out of 586 read, leaving 13 unmapped.
  • Partial correction: The description's "~190x cheaper per page" is right: $0.06810 per Opus page divided by $0.00036024 per Jev page is 189.0. Its "the full Opus pass would have run $43" is not reachable from the same numbers, which give $39.90.
  • Partial correction: "Co-inventor of ChatGPT" is press shorthand. The checkable fact is that Diogo Almeida (spelled Diogo, not Diego) is the fourth-listed author of arXiv:2203.02155, the InstructGPT paper, submitted March 4, 2022.
  • Verified: Jev's displayed $0.2111 reconciles with TypeSafe's published $0.042 per million input tokens and free output, implying 5.026 million input tokens, or 88.4 per judgment across 56,844 judgments. Both judgments-per-second columns match their own counters exactly (1,260.4 and 40.1).
  • Unverified: The Jev JUDGMENTS counter ends at 56,844, which is 6.47 times the 8,790 page pairs named in both the dashboard header and the description. I could not determine which quantity the cost figure is priced against, and the implied per-call token budget swings between 88 and 572 depending on the answer.
  • Clarification: Jev is proprietary. TypeSafe sells it as a hosted API and has released no weights or training data. The repositories named openjev and similar are third-party reimplementations of the interface on open base models, under Apache-2.0 (razorback16/openjev) or MIT (the rest).
  • Clarification: The run predates Jev 1.14. TypeSafe's docs date jev-1.14.0 to October 14, 2026 and still point jev-latest at 1.13.0, so a 2026-09-20 video can only show 1.13.
  • Context: Claude Opus 5 is now marked Legacy, released July 24, 2026, at $5 input and $25 output per million tokens with adaptive thinking at default effort high. Anthropic's lineup page routes "classification, extraction, and routing" to Claude Haiku 5.5, priced from $0.10 input per million tokens, which is one fiftieth of the Opus input rate the comparison uses.
  • Context: The Opus column's $1.43 leaves roughly 158 tokens per judgment if spent entirely on input. Even 200 output tokens per judgment would cost $9.05 at the documented $25 per million, so the shown spend implies near-zero output from a model whose default is adaptive thinking at high effort.
  • Context: RFC 9309 defines no Crawl-delay directive, and Google's robots.txt spec lists it among the fields it ignores. The live distribb.io/robots.txt sets none either, so no robots.txt rule bears on fetching 586 pages at 13 requests per second from that host. The sitemap at distribb.io/sitemap.xml currently lists 696 URLs.
  • Context: The description discloses that "the run stops the moment Jev finishes so Opus stops burning tokens", so Opus's 21 pages is a time-box cutoff rather than a failure point. The spoken version at 14.76 s states the time box correctly.
  • Context: No terminal output, API response, log file or crawler trace appears in any frame. Every figure in the video, including the 45.1 seconds and the 586 pages, originates from one dashboard built and rendered by Distribb, which is the creator's own product and the destination of the plugin he is promoting at 23.30 s.

Resources

Published September 20, 2026. Writeup generated from a favorited TikTok.