---
title: "Jev vs Claude Opus 5.5 for Email Sorting: Checking the 408x Claim"
date: "2026-09-25"
video_id: "7689528230956502286"
tiktok_url: "https://www.tiktok.com/@ccstrategic.io/video/7689528230956502286"
creator: "@ccstrategic.io"
category: "AI"
tags: ["jev", "claude opus 5.5", "openrouter", "email triage", "claude code"]
excerpt: "A TikTok says TypeSafe's Jev sorted 216 emails 408x cheaper than Claude Opus. At Anthropic's published rates the gap is closer to 256x. What else holds up."
---
Jev is a small decision model from TypeSafe, served through OpenRouter, that returns a label plus a confidence score instead of free text. A TikTok from @ccstrategic.io (Charles Dove, "Charlie Automates") says Jev sorted 216 emails 408 times cheaper and four times faster than Claude Opus. Jev's cost on screen is exact, but the $4.73 Opus figure does not match Anthropic's published rates, which put the same token counts at $2.98, about 256 times Jev's cost. The video also leaves out the accuracy numbers that justify the whole design.

## What is Jev?

Jev is built by TypeSafe, which opened early access to the model on 2026-09-15. It runs on OpenRouter as `typesafe/jev-1.13` and `~typesafe/jev-latest`. [OpenRouter's Jev guide](https://openrouter.ai/docs/guides/community/jev) describes a 32,000-token context window and three decision primitives: `choice`, `noul` and `score`. Each returns probabilities and a confidence value. `noul` is the real name of the yes-or-no primitive.

OpenRouter's [write-up](https://openrouter.ai/blog/insights/what-is-jev/) gives the price as $0.042 per million input tokens, with output free.

## Where to get the Claude x Jev skill

The video asks viewers three times to comment "JEV" to get the repo. The repo is public at [charlesdove977/claude-x-jev](https://github.com/charlesdove977/claude-x-jev), MIT licensed, created on 2026-09-25. The install line appears on screen:

```
npx claude-x-jev install --with-commands
```

The npm package `claude-x-jev` is at version 0.1.1. It adds slash commands to Claude Code, including `/jev-setup`, `/jev-bench` for running the benchmark on your own data, and `/jev-gate`, a tool-call safety hook.

## Is Jev really 408x cheaper than Claude Opus?

The Jev side checks out. The slide says 277,000 input tokens. At $0.042 per million, that is $0.011634, and the slide shows $0.0116.

The Opus side does not. The README names Claude Opus 5.5 as the comparison model and shows $4.73. Anthropic's [pricing page](https://platform.claude.com/docs/en/about-claude/pricing) lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens. Using the token counts on the same slide:

- 625,000 input tokens × $4 per million = $2.50
- 24,000 output tokens × $20 per million = $0.48
- Total: $2.98

That makes Jev about 256 times cheaper, not 408. Other Opus rates do not produce $4.73 either. Claude Opus 5 at $5 and $25 gives $3.73, and the retired Opus 4.1 at $15 and $75 gives $11.18. Reaching $4.73 at Opus 5.5 rates would take about 111,500 output tokens, roughly 4.6 times what the slide reports. Extra billed tokens, such as thinking tokens, cache writes or retries, would explain the gap, but the video does not say. The benchmark page is a local file, so the raw run cannot be checked.

Jev is still two orders of magnitude cheaper per decision for this kind of sorting.

## Is Jev four times faster than Claude?

Per email, yes. The screen shows Jev at 0.3 seconds per email and Claude at 1.3 seconds. The wall-clock row directly below tells a different story:

- Jev, 216 emails one after another: 65 seconds
- Claude, 9 batches at once: 37 seconds
- Claude, one after another: 4 minutes 46 seconds

Batched Claude finished the inbox faster than serial Jev. The README notes Jev takes "about 6 s in parallel." The video compares Jev's serial per-item time with Claude's batched per-item time.

## How accurate is Jev compared to Claude?

The video never says. The README does:

> When Jev said it was at least 70 percent sure (82 percent of the inbox), it matched Claude 95.5 percent of the time. When it was less sure, only 47 percent.

Overall agreement between the two was 87 to 94 percent. This is why the workflow uses a 70 percent cutoff. Of 216 emails, 178 route automatically on Jev's label and 38 go to Claude, cutting the combined cost to about $0.85 against $4.73 for Claude alone. At 95.5 percent agreement, roughly 8 of those 178 emails get a different label than Claude would give. Whether that is acceptable depends on what each label triggers.

The README also lists Jev's limits. It cannot follow a multi-condition playbook, has no memory between requests, and answers each question alone. In one case that produced an email tagged `vendor_pitch` and `affiliate_only` while also scored "92% likely to pay."

## Does using Jev save your Claude usage limit?

Yes. The cost slide mentions Claude's five-hour usage window. Anthropic's [usage limit docs](https://support.claude.com/en/articles/9797557-usage-limit-best-practices) confirm a rolling five-hour session limit plus weekly caps on paid plans. Jev calls are billed through OpenRouter and do not count against that pool.

## Other details in the demo

The terminal shows Claude Code running Fable 5.1 at xhigh effort, launched with `--dangerously-skip-permissions`. Anthropic's [models overview](https://platform.claude.com/docs/en/about-claude/models/overview) lists Opus 5.5 as current, with Opus 5 now legacy, so the comparison model was current.

Running `/jev-gate` in a session with permissions disabled skips the README's own advice. It recommends starting the gate in shadow mode with `JEV_GATE_APPROVE=1.01`, so nothing auto-approves, and moving to 0.9 only after reading the logs.

## Bottom line

Jev is far cheaper than Claude Opus for label-and-route decisions, at about 256 times by Anthropic's published rates rather than 408. Its speed lead disappears once Claude runs in batches. Route only Jev's confident answers and send the rest to Claude, as the README's accuracy numbers suggest.

## Sources

- [charlesdove977/claude-x-jev](https://github.com/charlesdove977/claude-x-jev) - the repo, install command and accuracy figures.
- [What Is Jev? (OpenRouter)](https://openrouter.ai/blog/insights/what-is-jev/) - pricing, builder and launch date.
- [OpenRouter Jev guide](https://openrouter.ai/docs/guides/community/jev) - the choice, noul and score primitives.
- [Anthropic pricing](https://platform.claude.com/docs/en/about-claude/pricing) - Opus 5.5 rates.
- [Claude models overview](https://platform.claude.com/docs/en/about-claude/models/overview) - current and legacy models.
- [Usage limit best practices](https://support.claude.com/en/articles/9797557-usage-limit-best-practices) - the five-hour window.
- [npm dist-tags for claude-x-jev](https://registry.npmjs.org/-/package/claude-x-jev/dist-tags) - the published version.
- [typesafe.ai](https://typesafe.ai) - the lab behind Jev.


## Video description and verification notes

**A 96-second vertical clip at 1080x1920, posted 2026-09-25 by @ccstrategic.io (channel name "Charlie Automates", real name Charles J Dove), sitting at 12,900 views, 441 likes, 34 comments, 130 reposts and 436 saves at capture, over an original sound credited to SAN577. I read all 48 frames and the full 389-word transcript. The top half of the screen cycles through four real artifacts: the GitHub README for a repo titled "Claude x Jev" with badges reading `npm v0.1.0`, `downloads package not found or too new`, `license MIT` and `stars 0`; a benchmark page loaded from a local file path (`/Users/user/Desktop/...`, tab title "Jev vs Claude · Brand-deal in...") showing `COST FOR ALL 216 EMAILS`, `Jev · $0.0116` against `$4.73 Claude Opus`, `277,000 input tokens. Output tokens are free`, `625,000 input tokens plus 24,000 output tokens at API prices`, and `Jev was 408x cheaper`; a five-step workflow card reading `1 Gmail / 216 emails`, `2 Jev sorts all / ~6 s · 1¢`, `3 Split on confidence / 70%+ sure: route by label. Under 70%: send to Claude / 178 auto · 38 to Claude`, `4 Claude thinks / ~35 s · ~$0.85`, `5 Charles approves`, with `COMBO COST ≈ $0.85 vs $4.73 for Claude alone` and `CLAUDE READS 38 emails instead of 216`; and a macOS Terminal (`Last login: Fri Sep 25 13:09:24 on ttys002`) running `claude --dangerously-skip-permissions`, banner `Claude Code v2.1.282 / Fable 5.1 with xhigh effort`, where he types `/jev-setup` from a slash-command menu that also lists `/web-setup`, `/seo-project-setup`, `/auto-mode-setup` and `/claude-x-jev`. The bottom half is a talking head under a purple "CHARLIE AUTOMATES" neon sign, with burned-in word-by-word captions in red and white.**

## The repo is public, named, and installable without commenting anything

The clip gates the deliverable behind a comment three separate times.

> And if you want access to this skill, just comment JEV. I'll send you the repo, but this is how you can set it up.

The repo is [charlesdove977/claude-x-jev](https://github.com/charlesdove977/claude-x-jev) on GitHub. It is MIT licensed, was created 2026-09-25 at 15:13 UTC, and had 19 stars and 2 forks when I checked on 2026-09-27. Frame 46 shows the install line on screen in plain text: `npx claude-x-jev install --with-commands`. The [npm registry](https://registry.npmjs.org/-/package/claude-x-jev/dist-tags) lists `claude-x-jev` at version 0.1.1 with 308 downloads in the 30 days to 2026-09-26. Anyone who reads the frame or searches the package name gets the same thing a commenter gets.

The "playbook" is less concrete. The README links to write-ups at [charlieautomates.com](https://www.charlieautomates.com/) and a free Skool community at `start.ccstrategic.io/skool`. The Skool URL resolves but returned HTTP 403 to my unauthenticated fetch, so I could not confirm what is behind it. The comment mechanic buys the creator engagement signal and a DM thread; it does not buy the viewer anything the README does not already hand over.

## The Jev side of the cost math is exact

OpenRouter's own write-up on the model gives Jev's price as [$0.042 per million input tokens with output free](https://openrouter.ai/blog/insights/what-is-jev/). Run the on-screen input count against that rate: 277,000 × $0.042 / 1,000,000 = $0.011634. The slide reads `$0.0116`. That traces to the cent.

The rest of the model description on screen holds up too. Jev is built by TypeSafe, an AI lab that entered early access with the model on 2026-09-15, and it serves through OpenRouter as `typesafe/jev-1.13` and `~typesafe/jev-latest`. [OpenRouter's Jev guide](https://openrouter.ai/docs/guides/community/jev) confirms the 32,000-token context window and the three decision primitives shown in the README table: choice, noul and score, each returning probabilities and a confidence value. One detail worth flagging for anyone squinting at frames 9 through 24: `noul` is the actual name of the boolean primitive. It is not an OCR artifact or a misspelling of `bool`.

## The Opus side does not reconcile with Anthropic's published rates

The README names the comparison model outright:

| | Jev 1.13 | Claude Opus 5.5 |
| --- | --- | --- |
| Cost | $0.01 | $4.73 |

Anthropic's [pricing page](https://platform.claude.com/docs/en/about-claude/pricing) puts Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens. Apply that to the token counts printed on the same slide as the $4.73:

- 625,000 input × $4 / 1M = $2.50
- 24,000 output × $20 / 1M = $0.48
- Total: **$2.98**

That is a 256x ratio against Jev's $0.0116, not 408x. No current or legacy Opus rate produces $4.73 from those counts either. Claude Opus 5 at $5/$25 gives $3.73. The retired Claude Opus 4.1 at $15/$75 gives $11.18. To reach $4.73 at Opus 5.5 rates you would need roughly 111,500 output tokens, about 4.6 times the 24,000 the slide reports. Something in that run was billed that the slide does not show, most likely thinking tokens, cache writes, or retries. The creator does not say, and the benchmark page is a local HTML file rather than a published artifact, so the raw run is not independently checkable. The README does ship a `/jev-bench` command to reproduce it on your own data, which is the honest part of the setup.

The direction of the claim survives. Jev really is two orders of magnitude cheaper per decision for this class of work. The specific 408x number does not survive contact with the price sheet.

## "Four times as fast" compares Jev serial against Claude batched

> And in this case, I found out it was four times as fast as Opus.

The per-email numbers on frame 13 back this: Jev at 0.3 s per email, Claude at 1.3 s per email, roughly 4.3x. What the spoken line drops is the wall-clock row sitting directly underneath it in the same screenshot:

- `Jev · 216 one after another`: 65 s
- `Claude · 9 batches at once`: 37 s
- `Claude · one after another`: 4 min 46 s

Batched, Claude finished the whole inbox in 37 seconds against Jev's 65 seconds run serially. The README's own table concedes Jev needs parallelism to win on wall clock, quoting "about 6 s in parallel." The clip picks the per-item serial figure for Jev and the per-item batched figure for Claude and calls the result 4x. Both numbers are real. The comparison is not apples to apples.

## The accuracy number that makes the whole design work never appears in the clip

The 70 percent confidence threshold on the workflow slide is the load-bearing part of this architecture, and the video never says why 70. The README does:

> When Jev said it was at least 70 percent sure (82 percent of the inbox), it matched Claude 95.5 percent of the time. When it was less sure, only 47 percent.

Overall agreement between the two models was 87 to 94 percent. That is the number a viewer needs to decide whether to let 178 of 216 emails route on Jev's label without a human or an LLM reading them. A 95.5 percent match on the confident slice means roughly 8 of those 178 emails get the wrong label. Whether that is acceptable depends entirely on what the labels trigger. The README is candid about the failure modes too: Jev cannot follow a multi-condition playbook, has no memory across requests, and answers each question in isolation, which produced at least one email simultaneously tagged `vendor_pitch`, `affiliate_only` and "92% likely to pay."

The 60-second format cut the one statistic that justifies the design. That is the clip's biggest omission, larger than the cost arithmetic.

## Everything else on screen is current as of 2026-09-27

Two days after upload, nothing in the clip has gone stale.

- Claude Opus 5.5 is the current default recommendation on [Anthropic's models overview](https://platform.claude.com/docs/en/about-claude/models/overview), with Claude Fable 5.1 above it for demanding reasoning. Claude Opus 5 is now listed as legacy. The README's choice of Opus 5.5 as the comparison model was correct at benchmark time and still is.
- The banner in his own terminal reads `Fable 5.1 with xhigh effort`, so he ran the demo on the current top-end model while benchmarking against Opus 5.5.
- The `5-hour usage window` phrasing on the cost slide is accurate. Anthropic's [usage limit documentation](https://support.claude.com/en/articles/9797557-usage-limit-best-practices) confirms a rolling five-hour session limit plus weekly caps on paid plans, and the point that API-metered Jev calls never touch that pool is correct.
- Claude Code `v2.1.282` and the `--dangerously-skip-permissions` launch flag are visible in frames 41 and 42. I could not verify that specific patch version against a published changelog.

The one thing worth noting about that terminal: he is demonstrating `/jev-gate`, a tool-call safety hook, inside a session launched with permissions disabled. The README's own guidance is to start the gate in shadow mode with `JEV_GATE_APPROVE=1.01` so nothing auto-approves, then go live at 0.9 after reading the logs. That caution does not make it into the video.

## Key Takeaways

- The free repo is [github.com/charlesdove977/claude-x-jev](https://github.com/charlesdove977/claude-x-jev), MIT, created 2026-09-25, 19 stars as of 2026-09-27. Install with `npx claude-x-jev install --with-commands`. Commenting "Jev" gets you nothing the README does not.
- Jev's cost on screen is exact: 277,000 input tokens at OpenRouter's published $0.042/MTok is $0.011634, shown as $0.0116.
- The $4.73 Opus figure does not reconcile. At Claude Opus 5.5's published $4/$20 per MTok, the 625,000 input and 24,000 output tokens on the same slide come to $2.98, making the real ratio about 256x rather than 408x.
- "Four times as fast" holds per email (0.3 s vs 1.3 s) and reverses on wall clock. The creator's own table shows Claude finishing 216 emails in 37 s across 9 parallel batches against Jev's 65 s run serially.
- The clip omits the accuracy data entirely. The README reports 87 to 94 percent agreement overall and 95.5 percent when Jev is at least 70 percent confident, dropping to 47 percent below that line. That single figure is what justifies the 70 percent routing threshold shown in step 3.
- `noul` is the genuine name of Jev's boolean primitive, confirmed in OpenRouter's Jev guide, alongside `choice` and `score`.
- The 5-hour Claude usage window referenced on the cost slide is real and documented by Anthropic. Metered API calls to Jev do not draw from it.
- **Unverified:** I could not confirm the raw 216-email benchmark, because the report is a local HTML file on the creator's Desktop rather than a published page, and the emails are private. I could not confirm how $4.73 was computed or which extra billed tokens account for the $1.75 gap. I could not confirm Claude Code v2.1.282 against a published changelog. I could not read `start.ccstrategic.io/skool`, which returned HTTP 403 to an unauthenticated request, so the contents of the "playbook" behind the Skool signup remain unchecked. I could not retrieve `openrouter.ai/typesafe/jev-1.13` directly (HTTP 404 to my fetch), so Jev's price comes from OpenRouter's own blog and guide rather than the model page.

## Resources

- [charlesdove977/claude-x-jev on GitHub](https://github.com/charlesdove977/claude-x-jev) — the actual repo behind the "comment Jev" gate. Confirms the install command, the 13 slash commands, the MIT license, the 2026-09-25 creation date, the Opus 5.5 comparison model, and the 87 to 94 percent agreement figures the clip omits.
- [What Is Jev? TypeSafe's Decision Model Explained (OpenRouter)](https://openrouter.ai/blog/insights/what-is-jev/) — confirms $0.042 per million input tokens, free output, TypeSafe as the builder, 32k context, and the 2026-09-15 early-access date.
- [OpenRouter Jev guide](https://openrouter.ai/docs/guides/community/jev) — confirms the `choice`, `noul` and `score` primitives and that each returns a confidence value.
- [Anthropic pricing, retrieved 2026-09-27](https://platform.claude.com/docs/en/about-claude/pricing) — confirms Claude Opus 5.5 at $4/MTok input and $20/MTok output, the rates that put the benchmark at $2.98 rather than $4.73.
- [Anthropic models overview, retrieved 2026-09-27](https://platform.claude.com/docs/en/about-claude/models/overview) — confirms Claude Opus 5.5 is current and Claude Opus 5 is legacy, so the clip's model choice is not out of date.
- [Anthropic usage limit best practices](https://support.claude.com/en/articles/9797557-usage-limit-best-practices) — confirms the rolling five-hour session window and weekly caps referenced on the cost slide.
- [npm dist-tags for claude-x-jev](https://registry.npmjs.org/-/package/claude-x-jev/dist-tags) — confirms the package is published at v0.1.1, one version past the v0.1.0 badge visible in the video.
- [typesafe.ai](https://typesafe.ai) — the lab that builds Jev, resolves and live as of 2026-09-27.
