model fatıgue

Corrections

Every change to a video or its page after publication is listed here, newest first.

A correction means we got something wrong. A source update means someone else revised their number after we used it, and the video was right when it was made. A retest means we ran the same test again on a new build. A reprice means a price changed and the measurement didn't.

On YouTube, a correction gets a "Correction:" line in the description and a pinned comment. A source update gets a pinned comment. The page shows the old and the new value side by side, with dates.

12 Oct 2026 · reprice

Does JevBench's new #1 really beat Jev?

We said: Nobody we could find sells Quyet as a service, so it has no price, and for the smallest bill, Jev.

Now: That was true when the page went up on 7 October. FPT AI has sold Quyet-1.0-Large as a hosted API since later that day, at $0.08 per million input tokens. At that price, with the tokens of our run, Quyet would cost about $0.022 per thousand decisions, a little less than Jev's $0.026. We haven't run it there. The page now says so.

7 Oct 2026 · source update

Does JevBench's new #1 really beat Jev?

We said: JevBench v1.6.1 now tests hosted APIs such as Jev on the full question set without adjusting their scores, its official ranking is now Sage 1.3.0 first, Jev second and Quyet third, Sage is now second on Capability, and Clef-Flash is number 12.

Now: Those are v1.6.1 as we read it on 6 October, and the page now says so. JevBench has since split its live leaderboard into open models and hosted APIs and added twelve open models to v1.6.1, which moved some places, Clef-Flash's among them (to 15th among the Jev-class systems and 24th officially, as read on 7 October), but no scores.

On YouTube: The description was updated the same way on 7 October, and a pinned comment explains the change.