A Chinese model prices like an American one for the first time
Kimi K3's API pricing: per million tokens, $0.30 cache-hit input, $3 cache-miss input, $15 output.
The last two — $3 input, $15 output — are identical to Anthropic's Claude Sonnet 5.
This needs underscoring: Chinese open-source models used to win on rock-bottom pricing. This time Moonshot priced straight into the US tier. Not winning on cheap — pricing on capability.
But against its own predecessor? K3's API input is up over 3x and output nearly 4x versus K2.6. Whether the market accepts that hike comes down to retention.
US markets voted first
The day after the news, US tech stocks took a beating. The Nasdaq composite dropped 1.4%, semiconductors fell harder.
What the market is repricing: if the best open model can wrestle the best closed models to a draw — or outright beat them in a frontend coding blind test — what justifies what closed vendors charge?
Former Meta senior product manager Qu Xiaoyin put it bluntly: when the best open model surpasses the best closed one, how does Anthropic defend Fable's pricing? Who's still paying that?
She added a starker math problem: two months ago Moonshot was valued at $20 billion, Anthropic near $1 trillion — 50x the gap. A 50x valuation gap, but the product gap is laid bare on the leaderboard.
Open costs a sixth of closed
This comparison bites harder: open-model usage costs average a sixth of closed competitors, while the performance gap keeps closing.
The CFO's old dilemma was "is 1-3% more performance worth paying 6x?" Now that both the performance edge and the cost edge sit on the open side, that math doesn't pencil anymore.
What Moonshot is betting on
Zoom out from the model to the company. Moonshot is prepping a Hong Kong IPO, and K3 landing at this moment isn't just a product refresh — it reads like a capability flex ahead of listing.
What backs the valuation is revenue velocity:
- March: ARR crossed $100 million
- Mid-June: ARR hit $300 million, tripled in three months
- API accounts for 70%+ of revenue, and overseas API now exceeds domestic
The subscription data is wilder. Per Stripe, Kimi's personal-subscription paid orders jumped 8280% month-over-month in January, then another 123.8% in February — cracking Stripe's global top 10, the first Chinese AGI product to do so.
Can the open-source flywheel actually spin
Moonshot's playbook: use the open model to pull in global developers, get into real workflows via dev tools, then convert users who need stability and speed into API customers.
In theory it's a flywheel: model capability builds open-source influence, that brings developers, developers convert to API revenue, revenue funds the next round of training and inference.
Whether it actually spins comes down to a few numbers: developer retention, API conversion rate, and whether inference cost can be controlled. K3's 4x price hike and whether growth holds is the first real test.
Traditional stats are going blurry
Here's a telling mismatch: on July 16, IDC published its 2025 China AI coding market share report — total size 399 million RMB, with Alibaba's Qoder first at 47.6%.
399 million RMB? Moonshot alone has $300M in ARR. The numbers don't line up — not because someone miscounted, but because the yardstick changed.
Traditional software sells licenses; that's what IDC measures. Model companies earn through API calls and subscriptions, which the old framework can't capture. The value in AI coding is shifting from software sales toward API and subscription.
The real shift
K3's biggest signal isn't "open source caught up another step."
It's that open models started pricing on tier. Open source used to mean "cheap but a bit worse." Now it dares to match closed flagships on price — and wins blind tests. When price is no longer open source's only selling point, the foundation of the closed moat loosens.
Eighteen months ago, the question for Chinese models was "can they catch GPT-4." Now the benchmark is Fable 5 and GPT-5.6 Sol. K3 may not be the one that ultimately wins, but it proved one thing: the gap has compressed to within a single leaderboard.
Whether this US stock drop is overreaction or fair repricing — that depends on whether closed vendors' API revenue holds up next quarter.