Home › Q&A › Kimi K3 lists at $3/$15 per million toke
Kimi K3 lists at $3/$15 per million tokens: can switching to it rescue an AI app's gross margin, and what are the catches?
Asked by：创始人·AI 应用 CEO / CTO · Judged · Ledger updated · 张倩 Cynthia Zhang · FutureX Capital
Yes, but the diary's sample, Harvey, got there by post-training on a Kimi K3 base (the company's own statement), not by only repointing the API. Its gross margin went from about 50% early in the year to -50% in June and turned positive after it released its own model in August, per Bloomberg on 9/21 citing people familiar; Harvey declined to comment.
Judgment
The mainstream view is that if money moves, it runs to the cheapest model. The diary's reading is that buyers on the Vercel gateway stepped down only one price tier, and Kimi K3 at $3/$15 is a mid-tier price. The weights are open, but pricing power stays in the license and the official API, so the savings come with conditions.
Facts
Vercel model page, 9/29: Kimi K3 lowest price $1.325/$10.50, while most providers still charge $3/$15. Vercel export (a list-price-weighted estimate, not gateway billing): its spend share averaged 9.3% over the 30 days from 8/30 to 9/28 and peaked at 18.5% on 9/22, a pulsed pattern. Moonshot AI raised its official API output price from ¥27 to ¥100 per million tokens 11 days before the weights went live, about 3.7x (ENTRY 114 listed no source line; not independently recovered).
What to do
For founders: move agent workloads first, compute margin twice (at the gateway's lowest price and at the raised official price), open talks on the license revenue gate a year ahead, and state in customer contracts who generates the reply and where data goes. For investors: in diligence ask how much of the company's value disappears if the base model is swapped, and among Chinese models back the ones that can collect money above $10/M output.
Falsifier (verbatim from the diary footnote)
The two claims are ruled separately. The ruling source for both is the official Vercel labs export recomputed with the repository script scripts/vercel-lab-share.py, readings public; whichever triggers first decides; a 30-day window may not end before the start date, and 9/29 is counted at its complete value from a later re-pull. (2) "Kimi K3 collecting money at a mid-tier price is not a batch pulse": starting 2026-09-29, through 2026-11-30, if Kimi K3's spend share over any consecutive 30 days averages below 8% (the 30-day average for 8/30 to 9/28 was 9.3%), mark ✗; if any consecutive 30-day average is not lower than 12%, mark verified; if neither triggers by the deadline, mark unruled and close; ⏳ in the meantime. Event branches: if before 2026-12-31 a second US company valued at no less than $1 billion confirms, by company announcement or Bloomberg/Reuters reporting, that its production model is based on Kimi K3, (2) is upgraded; if before 2027-03-31 the US Treasury or Commerce Department sanctions Moonshot AI, or AWS/Vercel delist Kimi K3 (official announcement governs), the overseas money-collection chain in (2) is void and is merged with falsification clause (1) of ENTRY 135.
Disclosure: This diary is written and verified with Claude; discount any judgment about Anthropic, favourable or not; FutureX Capital has fully exited Meituan (confirmed by the author on Oct 10, 2026), so it no longer has an indirect interest in Moonshot AI through Meituan; The author is a Vercel customer; Vercel's leaderboard 'spend' is a list-price-weighted estimate, not billing.