Jev after the promotion: rate ambiguity and a new comparison
Selected public sources were checked about 05:03–05:07 UTC. This is a dated change audit, not an account test or a rerun of any benchmark. The exact 24-hour monitor covers 25 September 05:03:32 through 26 September 05:03:32 UTC.
Route changes, not account quotations
| Route | 26 September public observation | Unverified |
|---|---|---|
| TypeSafe direct | jev-1.13.0; $0.042/M input tokens, output free. jev-latest and jev-preview pointed to 1.13.0; listed limits are dynamic. | Account access, quota and bill. |
| OpenRouter | typesafe/jev-1.13 through its decisions API; $0.042/M input, $0 output. | Credits, alpha-route access and bill. |
| Cloudflare Workers AI | The Jev-specific page now lists $0.042/M input, $0 output, $0 cached input, 32k context and “Zero data retention: Yes.” This replaces our earlier dashboard-only price note. | Account quota, bill and the data path beyond this route; its ZDR field is not an application-wide promise. |
| Vercel AI Gateway | Free remains displayed for TypeSafe AI and DigitalOcean rows after the stated 25 Sep promotion end. | Effective Jev rate and eligibility. DigitalOcean is a provider row inside Vercel, not a separately verified direct route. |
| Netlify AI Gateway | TypeSafe SDK in Functions, billed through Netlify credits; no Jev-specific USD/token conversion published here. | Credit conversion, route entitlement and bill. |
| Opper Jev | Opper documents POST https://api.opper.ai/v3/compat/v1/systemone with typesafe/jev-latest. Its card rounds to $0.04/M input; the article claims $0.042/M. Neither is a measured bill. Card says ZDR not established. | Entitlement, actual routing, region, terms and billed rate. |
One operator-reported Jev–Kev suite
Opper’s 25 September article, code, 362-item manifest and results are primary artifacts. Opper says it sent identical one-question bodies through its own System One endpoint to Jev 1.13 and Kev 4B, with eight calls in flight. Items were 160 arXiv categories, 120 Stack Exchange sites and 82 GitHub issue labels. One run each, on tasks where both were about 94–98% accurate, does not establish general superiority. Opper reports Jev lower 10-bin calibration error on all three tasks, about 257 more counted input tokens per Jev request, and 220 versus 275 ms concurrent median. Its very short 280-versus-23-token example is not a workload-wide 12× ratio; packing may amortize fixed request cost. Network location and route affect latency. Opper hosts one compared model and sells route access. We did not rerun its scripts or verify raw provider usage or a bill.
Artifact and product boundaries
The convaiinnovations/laya card directly links NandhaKishorM/laya, whose quick start loads that model. Both show Apache-2.0. The separate he-jev/laya references them, but no checksum or provenance declaration established identical files or maintenance. The model card’s own root-checkpoint limitations must not be hidden behind a fine-tuned score.
The Qwen3.8-27B base card lists Apache-2.0. JevAny claims two rank-16 adapters, but its SFT and RLCR cards were inaccessible: adapter file identity and individual licence remain unverified. Base licence does not transfer automatically.
Jevmail documents a Vercel call but its no-key preview is simulated; its “free” cost language cannot resolve the current Vercel contradiction. DocJev has a small author-run 19 September document comparison, not a general performance guarantee. WordPress Jev Comment Triage documents a call path, not a verified production installation. Jev Trade defaults to mock, testnet and simulated fills without a key; its public desk showed Offline and blank positions/trades. No live Jev trade, fill or mainnet operation was verified.
Preceding-24-hour monitor and limits
A creator Reddit post displayed about 40 minutes of age at an approximately 05:05 UTC check on 26 September, so its announcement is reasonably in-window; that is relative-time evidence, not an exact timestamp. Opper’s article and repository have 25 September dates without hours and are date-ambiguous discoveries. Laya metadata says last modified 24 September 05:39:22 UTC, outside this window. GitHub topic updates were date-only, and exact created/pushed records could not be read. Native X search returned 403; TikTok, YouTube, Reddit new/day, GitHub search/history and HN Algolia were inaccessible or stale. No verified item on a surface is not proof the surface was empty.
All dated reports · Current route table · Alternatives by layer.