TL;DR
Moonshot AI released Kimi K3 on July 16 at $3 per million input tokens and $15 per million output tokens, matching Claude Sonnet 5’s list price and costing about five times more than its predecessor. Independent testing placed K3 near the model frontier, but its weights, licence, technical report and active parameter count remain unavailable.
Moonshot AI released Kimi K3 on July 16 at $3 per million input tokens and $15 per million output tokens, matching the list price of Anthropic’s Claude Sonnet 5 and costing about five times more than its predecessor. The launch matters because an independent evaluation placed K3 close to the leading model tier, while Moonshot’s pricing suggests it no longer sees a steep discount as necessary to compete.
K3 is available through the Kimi app, Playground and API. Moonshot describes it as a 2.8-trillion-parameter mixture-of-experts model that routes 16 of 896 experts per token, supports text, image and video input, and offers a maximum context window of 1,048,576 tokens. Only the Max reasoning setting was available at launch.
On the independent Artificial Analysis Intelligence Index v4.1, K3 scored 57.1, compared with 59.9 for Claude Fable 5 using an Opus 4.8 fallback and 58.9 for GPT-5.6 Sol Max. That left K3 2.8 points behind the highest score in the supplied results. Artificial Analysis also recorded a 732-point Elo increase over K2.6 on its long-horizon tracker, taking K3 to 1,547, and ranked it first on Design Arena.
The pricing marks a clear change from the K2 family’s approximate $0.60 input and $3 output rates. K3 also costs more during Claude Sonnet 5’s temporary introductory period: Anthropic’s model is listed at $2 input and $10 output through August 31, making K3 50% more expensive during that promotion.
Kimi K3: the gap closed six months early — and China stopped competing on price
Every write-up today says “China caught up.” True — and the less interesting half. The other half: K3 costs 5× its predecessor, making it the most expensive Chinese model ever, priced at exact parity with Claude Sonnet 5. A benchmark is a claim. A price is a claim the vendor has to live with.
For two years the thesis was “cheap alternative.” Moonshot just abandoned it. Vendors discount when they’re compensating for something — Moonshot has stopped compensating. With Sonnet 5’s intro rate at $2/$10 through 31 Aug, K3 currently costs 50% more than the model it’s priced against. The competition just moved from cheap vs good to good vs good at the same price, with one of them open — and you can’t answer that with a discount.
The story we’ve told: export controls forced Chinese labs into efficiency. But K3 is 2.8T — the largest open model ever, ~3× K2, vs DeepSeek V4-Pro’s 1.6T. That’s not more with less. That’s more with more. Caveat: sparse MoE, active params undisclosed — total ≠ FLOPs. But if the controls were binding at the frontier, this model shouldn’t exist.
Anthropic has accused Moonshot, Z.AI, MiniMax, Alibaba & DeepSeek of “illicit” distillation — possibly well-founded; I can’t assess it. But one day earlier, Thinking Machines said Inkling’s post-training bootstrapped on Kimi K2.5 — reported as ecosystem health. Same verb, different flag, different word. If the distinction is real, someone should articulate it.
Two things changed, neither in the headlines. The discount is gone — anyone whose China strategy was “they’re cheaper” needs a new strategy. And the controls didn’t work — six months early, biggest model ever, from a lab that was supposed to be compute-starved, while Washington’s options narrow to loosening restrictions on its own labs, criminalising distillation, or subsidising American open weights. That’s not containment. It’s a menu of concessions. The gap is 2.8 points and closing. The price is Sonnet’s. The weights are ten days out. Everything that matters happens on 27 July.
Moonshot Drops the China Discount
Chinese model developers have often been positioned as lower-cost alternatives to leading US systems. K3’s price weakens that distinction: Moonshot is asking customers to compare it with Claude Sonnet 5 on capability, deployment options and performance, rather than cost alone.
Thorsten Meyer AI argues that the launch moves competition from “cheap versus good” to “good versus good”. That is an interpretation, not a settled market outcome. Customers will still need evidence on reliability, latency, tool use and total operating cost before price parity can be treated as capability parity.
The GPT-4 Millionaire: Future of Business Featuring Microsoft 365 Copilot: How to Leverage AI Language Models to Grow Your Company and How AI-driven Language Models Will Revolutionize the Way We Work
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
K3 Arrives Ahead of Forecasts
The supplied analysis says models at K3’s performance tier had been expected in early 2027, leading to the claim that Moonshot closed the gap about six months early. No underlying analyst forecast was included, so that timetable cannot be independently assessed from the provided material.
Moonshot calls K3 its most capable model to date and says it has 2.8 trillion total parameters, around three times the scale of the K2 family. That scale complicates claims that Chinese labs have advanced only through efficiency under export restrictions. Still, K3 is a sparse mixture-of-experts system, and total parameters do not reveal how much computation each request uses.
“Our most capable model to date, with 2.8 trillion parameters.”
— Moonshot AI launch materials
Licence and Compute Details Missing
K3’s weights were not available at launch; Moonshot has promised them by July 27. The licence has not been published, so descriptions of K3 as open source or permissively open weight remain premature. Its technical report and active parameter count are also unpublished.
The announced one-million-token context window is a maximum, not a guarantee across every service tier; the Moderato tier is capped at 256,000 tokens. It is also unclear how K3 performs under sustained production workloads, whether Moonshot will retain its current prices, and how broadly the independent results will hold across coding, reasoning and agent tasks.
July 27 Becomes the Test
Attention now turns to Moonshot’s planned July 27 weight release. Developers will examine the licence, hardware requirements, active parameter count and reproducibility of the reported results. Those disclosures will determine whether K3 offers a practical open-weight alternative or remains primarily a hosted model competing at Western commercial prices.
Key Questions
What does Kimi K3 cost?
K3 costs $3 per million input tokens and $15 per million output tokens. Cached input is listed at $0.30 per million tokens.
Is Kimi K3 as capable as the leading Western models?
Artificial Analysis placed K3 near the leading tier, with a score of 57.1 and a 2.8-point gap to the top result supplied. That does not establish equal performance across all tasks or production settings.
Is Kimi K3 open source?
Not at the time of publication. Moonshot has promised model weights by July 27, but the licence and technical report remain unpublished.
Why is K3’s pricing receiving attention?
The price is about five times higher than the K2 family’s rate and matches Claude Sonnet 5’s standard list price. It suggests Moonshot wants K3 judged on capability rather than a large price discount.
Does the 2.8-trillion-parameter figure show K3’s computing cost?
No. K3 routes only 16 of 896 experts per token, and Moonshot has not disclosed the active parameter count. Total model size cannot by itself establish training or inference cost.
Source: Thorsten Meyer AI
