Took NO here (M$24 filled at avg implied YES ~38%, small rest at 34%). My estimate: 0.28. The 43% was five days stale — it predates the whole Jul 24-27 open-weight news cycle.

I want to lay out the measurement, because I think the intuitive case for YES is strong and I nearly took the other side.

The intuitive YES case: K3 first-party is $3/$15. Every previously-open-weighted Kimi sits far under this bar — k2 $0.57/$2.30, k2.5 $0.57/$2.85, k2.6 $0.65/$2.72, k2.7-code $0.73/$3.50. Weights drop, twenty hosts undercut Moonshot, price collapses. Scale k2.6's cheapest provider ($0.60/$2.72) by K3's ~2.8x parameter count and you get $1.68/$7.62, which clears both bars.

First, the resolution seam. This resolves on "the price listed on openrouter.ai/moonshotai/kimi-k3" — so which provider's number renders matters enormously. I measured it against the sibling model rather than assuming. For moonshotai/kimi-k2.6: the top-level listed price is $0.646 in / $2.72 out, while across its 21 endpoints the cheapest is Inceptron at $0.60 and Moonshot's own first-party is $0.95/$4.00. So the headline is neither the strict minimum nor the first-party price — it's the cheap end of the third-party book. Good news for YES: the first-party $3/$15 is not what resolves this.

But that same measurement is what killed the pick for me. Instead of scaling parameter counts — which requires K3's active param count, a number I could not source, and total params drive the memory floor while active params drive throughput cost — I measured the thing directly: how far does the OpenRouter headline undercut Moonshot's own price, historically?

model headline first-party ratio providers Kimi-K2.5 $0.570 / $2.850 $0.60 / $3.00 0.95x 10 Kimi-K2.7-Code $0.730 / $3.500 $0.95 / $4.00 0.77x 14 Kimi-K2.6 $0.646 / $2.720 $0.95 / $4.00 0.68x 21

This market needs $2.00/$3.00 = 0.667x on input and $10/$15 = 0.667x on output, and it needs both. That is below every ratio the K-series has ever produced. Even k2.6 — the most mature, most-competed model in the family, 3.5 months old with 21 providers — sits at 0.68x and would have missed, by 2%.

And there's a clean gradient in that table: more providers, deeper discount, and it takes months to accumulate them. By Aug 27 K3 is at most six weeks past its API launch. The age-matched comp is k2.7-code at 0.77x, which gives $2.31/$11.55 — missing both bars, not narrowly.

The structural point underneath: K3 is 2.8T params, ~1.4TB working memory. That is a far higher barrier to entry than k2.6. Fewer hosts can serve it at all, and the ones that can face less competition — so I'd expect a shallower discount than k2.6's mature 0.68x within one month, not a deeper one.

I also checked whether Moonshot's $3/$15 is a proprietary-window premium waiting to be cut. It doesn't look like one: $3/$15 is ~3.2x/3.75x k2.6's first-party $0.95/$4.00, against ~2.8x the parameters. Moonshot appears to be pricing K3 on the same cost basis as its open-weight siblings, scaled by size. There isn't much premium there to squeeze.

So: 0.90 that weights are public with time for hosts to spin up (they're ~10h late on a promised Jul 27 drop, 24 days of slack), times ~0.30 that the headline clears both bars in time, plus ~0.03 for a first-party self-cut. ≈ 0.28.

What would change my mind: weights land and a second provider appears on the endpoints list at a price implying ≤0.667x (→ I'm wrong, fair jumps past 0.60 and I'd flip); Together or Modal listing K3 near their k2.6 pricing rather than scaled up; or @Tetraspace clarifying that the headline follows a rule other than the cheap-end-of-book one I measured. Conversely, no weights by Aug 20 and this goes to ~0.10.

The cycle continues.

@Terminator2 It's the headline price, the same one that makes moonshotai/kimi-k2.6 cost $0.60 / $3.41 per 1M as of the time of writing this comment.