Kimi K3's first 48 hours, measured: 72B tokens a day at 5× the family price
Moonshot AI released Kimi K3 on July 15, 2026. TKX tracks it the only way we track anything: by measurement. Forty-eight hours in, the numbers are already unambiguous — K3 routed 72 billion tokens in a single day (July 17 data), instantly the busiest model in the Kimi family, at a list price five times its predecessor's.
The launch numbers
| Metric | Measured value | Basis |
|---|---|---|
| Routed volume (24h) | 72.0B tokens/day | OpenRouter daily usage dataset, 2026-07-17 |
| Est. spend (24h) | ≈$432K/day (est.) | paid tokens × list rate, 3:1 in:out blend |
| List price | $3 / $15 per 1M (in / out) | OpenRouter listing |
| Context window | 1,048,576 tokens | OpenRouter listing |
| Free-tier share of volume | 0% — all paid traffic | permaslug variant split |
Live row, refreshed hourly: kimi-k3 on the model board.
The pricing bet
K3 is not an iteration on K2's pricing — it is a repositioning. The family ladder, at list:
| Model | In / Out ($/1M) | Volume (24h, 07-17) |
|---|---|---|
| kimi-k3 | $3.00 / $15.00 | 72.0B |
| kimi-k2.5 | $0.57 / $2.85 | 53.2B |
| kimi-k2.6 | $0.95 / $4.00 | 51.2B |
| kimi-k2.7-code | $1.00 / $4.40 | 18.5B |
At $3/$15 with a million-token window, K3 is priced squarely against the frontier closed tier — the same list point as premium western flagships — rather than against the open-weight value tier where the K2 line lives. That a Chinese lab's model holds 72B tokens/day of fully-paid traffic at that price within 48 hours is the actual story: willingness-to-pay is the one signal launch benchmarks cannot fake.
Cannibalization? Not yet
K2.5 and K2.6 still carried ~53B and ~51B tokens respectively on the same day — K3's launch volume is so far additive, not substitutive. Whether the family total holds is exactly what a daily measured series is for; the curve will be visible on the board either way.
The honest limits
Methodology: how TKX measures · live data: usage API · how the usage board works.
← All insights