<- all tokdocs

Jev vs Claude Opus 5 for SEO Internal Linking: Checking the Demo

By , founder of Mavric and Apso

View on TikTok ->

Jev is a decision model from TypeSafe AI that takes text and questions and returns typed answers with probabilities, priced at $0.042 per million input tokens. A TikTok from @itsborjao says Jev "read through all 586 pages" of a website and built its internal link map in 45 seconds for 21 cents, while Claude Opus 5 handled 21 pages for $1.43. The per-page cost gap roughly checks out, but the dashboard labels itself a "simulated run" at "2x real" speed, and Jev does not fetch web pages.

What is Jev?

Jev is TypeSafe AI's first public model. Artificial Intelligence News reported that TypeSafe left stealth on September 16, 2026, with 70 to 500 ms end-to-end latency, $0.042 per million input tokens and no output charge. It quotes founder Diogo Almeida: "Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out."

TypeSafe's model reference lists Jev 1.14 as released on October 14, 2026, with jev-latest still pointing to 1.13.0. A video from September 20 could only have used 1.13.

Jev is a hosted API. TypeSafe has not released weights or training data. GitHub repositories named "openjev" and similar are third-party reimplementations of the interface on open models, not TypeSafe code.

Was Jev built by the co-inventor of ChatGPT?

The video calls Almeida "the co-inventor of ChatGPT." The checkable fact is that Diogo Almeida (not Diego) is the fourth-listed author of arXiv:2203.02155, "Training language models to follow instructions with human feedback," submitted March 4, 2022. That is the InstructGPT paper, the RLHF work ChatGPT was built on. "Co-inventor of ChatGPT" is press shorthand, not a title any primary source gives him.

Did Jev really read 586 pages in 45 seconds?

No. TypeSafe's docs say 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." It has no crawler.

The dashboard in the video agrees. Its header reads "586 pages crawled off distribb.io · 15 destinations · 8790 page pairs · simulated run." Distribb's tool crawled the pages beforehand. The 45.1-second clock covers the decision pass over pages already on disk.

The clock also runs fast. A speed selector on the dashboard reads "2x real." The Jev elapsed counter moves 43.6 seconds over 22.0 seconds of video, a ratio of 1.98, which is why a 45.1-second run fits in a 28-second clip. Treat the comparison as a replay of an earlier result, rendered by Distribb, the creator's own product. The video shows no terminal, API response or log.

Is Jev really 190x cheaper than Claude Opus 5?

Per page, roughly yes. The description says "per page that is ~190x cheaper." Jev's $0.2111 over 586 pages is $0.00036 per page. Opus's $1.43 over 21 pages is $0.0681 per page. The ratio is 189.

Jev's cost also matches its list price. $0.2111 at $0.042 per million tokens buys 5.026 million input tokens, or 88.4 tokens for each of the 56,844 judgments on screen.

Other numbers do not line up:

  • The description says "the full Opus pass would have run $43." At the dashboard's own Opus rate, 586 pages cost $39.90.
  • The description says "584 links placed" and "139 pages it refused to link." The dashboard shows 573 mapped out of 586 read, leaving 13 unmapped.
  • The header says 8,790 page pairs (586 pages times 15 destinations), but the Jev judgments counter ends at 56,844, about 6.47 times higher. The video does not explain the difference.

The Opus figure is also low for that model. Claude Opus 5, released July 24, 2026 and now marked Legacy, costs $5 per million input tokens and $25 per million output tokens, with adaptive thinking at a default effort of high. Even 200 output tokens per judgment across 1,809 judgments would cost $9.05, so the $1.43 implies almost no output.

Is Claude Opus 5 the right model to compare against?

Probably not. The description calls the job "8,790 yes/no calls ... that is a classification problem." Anthropic's models overview recommends Claude Haiku 5.5 "For high-volume, latency-sensitive tasks such as classification, extraction, and routing," from $0.10 per million input tokens. That is one fiftieth of Opus 5's input price. Anthropic's Batch API would halve the Opus cost again, to $2.50 input and $12.50 output.

The description also says "the run stops the moment Jev finishes so Opus stops burning tokens." Opus's 21 pages is where the clock stopped, not where Opus failed.

How does Jev hit 1,260 judgments per second?

The dashboard shows 1,261 judgments per second for Jev, which matches 56,844 divided by 45.1. TypeSafe's docs say rate limits are per organization, "measured in tokens per second and requests per second," without published numbers. They also say "There is no separate limit on the number of questions in a request," within a 64k context per request and 32k per question. Batching many questions into each request is the documented route to that throughput.

Does robots.txt limit crawl speed?

Not by any standard. RFC 9309, the Robots Exclusion Protocol, defines no Crawl-delay directive, and Google's robots.txt spec ignores it. The site in the demo, distribb.io/robots.txt, sets no crawl delay either. Its sitemap listed 696 URLs as of 2026-10-19.

Bottom line

Jev is a real, cheap, fast decision model, and its per-page cost in this demo is about 189 times lower than Claude Opus 5. The demo is a 2x-speed simulated replay, the crawl happened separately, and several numbers in the description do not match the dashboard. A fair comparison for a yes/no classification job would be Claude Haiku 5.5 or Opus on the Batch API.

Sources

Published September 20, 2026. Based on a TikTok by @itsborjao. Full video notes (markdown)