TL;DR
Moonshot AI released Kimi K3 on July 16, placing within 2.8 points of the leader in the Artificial Analysis Intelligence Index. Its Western-level API pricing marks a shift from competing mainly on cost, although claims that it arrived six months early rely on an analyst forecast cited by the source.
Moonshot AI released Kimi K3 on July 16, bringing the Chinese model within 2.8 points of the leading system in an independent Artificial Analysis test and pricing its API at $3 per million input tokens and $15 per million output tokens. The result narrows the measured performance gap while signaling that Moonshot intends to compete with Western providers on capability rather than price alone.
Kimi K3 scored 57.1 on Artificial Analysis Intelligence Index v4.1, compared with 59.9 for the highest-scoring configuration. The source material also reports that K3 placed first on Design Arena and recorded a 732-point Elo increase over Kimi K2.6 on the evaluator’s long-horizon tracker. Those figures are independent of Moonshot, but they reflect testing available only one day after release.
Moonshot describes K3 as a 2.8-trillion-parameter sparse mixture-of-experts model that routes 16 of 896 experts for each token. It supports text, image and video input and advertises a maximum context window of 1,048,576 tokens. The company has not disclosed the active parameter count, and only the Max reasoning setting was available at launch.
The model is live through the Kimi app, Playground and API. Its listed API rates are about five times those attributed to the K2 family and match Claude Sonnet 5’s standard $3/$15 pricing. Sonnet 5’s temporary introductory rates of $2/$10, scheduled through August 31, leave K3 about 50% more expensive during that period.
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
K3’s pricing challenges the established view of Chinese models as lower-cost substitutes for Western systems. Moonshot is asking customers to pay Western mid-tier rates, suggesting confidence that the model’s performance can support direct comparison on quality.
The competitive question now extends beyond benchmark scores. Buyers will need to compare reliability, latency, tool use and total operating costs at similar API prices. If Moonshot releases usable weights under permissive terms, K3 could also offer self-hosting options unavailable from many closed-model providers.
AI language model API
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A Forecast Beaten, Not a Deadline
The claim that K3 arrived six months ahead of schedule does not refer to a Moonshot deadline. The source analysis says analysts had expected a Chinese model to reach this performance tier in early 2027, but it does not identify a formal forecast or a common industry timetable.
K3 also complicates the argument that export restrictions have forced Chinese laboratories to rely mainly on efficiency. Its 2.8-trillion total parameter count is roughly three times that of Moonshot’s K2 family. Because it uses sparse routing and its active count remains undisclosed, total parameters cannot be treated as a direct measure of computing demand.
“Our most capable model to date, with 2.8 trillion parameters.”
— Moonshot AI launch materials
AI development tools for developers
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Weights and Licence Still Missing
K3 cannot yet be independently described as an available open-weight model. Moonshot has promised model weights by July 27, but the licence and technical report remain unpublished. Until those documents appear, users cannot verify redistribution rights, commercial restrictions or deployment requirements.
Moonshot’s own benchmark results remain self-reported, while the independent Artificial Analysis figures cover an early tested configuration. Real-world performance across coding, agents and long-context workloads is still developing. The release also does not establish that US export controls have failed; that conclusion would require evidence about K3’s training hardware, computing access and development costs.
large language model API access
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
July 27 Becomes the Test
Attention now turns to Moonshot’s planned July 27 weights release. The licence, technical report and hardware requirements will determine whether K3 can be independently reproduced, modified and deployed. Further third-party testing should also show whether its early benchmark standing holds across broader workloads.
AI model performance testing tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Did Kimi K3 fully close the gap with leading AI models?
No. K3 scored 57.1 in Artificial Analysis v4.1, leaving it 2.8 points behind the leader. The result indicates a narrow measured gap, not performance parity across every task.
Was Kimi K3 officially released six months early?
No formal Moonshot schedule is cited. The six-month claim compares the July release with an early-2027 analyst expectation referenced in the source analysis.
How much does the Kimi K3 API cost?
Moonshot lists K3 at $3 per million input tokens and $15 per million output tokens, with a reported cached-input rate of $0.30 per million tokens.
Is Kimi K3 open source now?
No. The weights were not available on July 17, and the licence had not been published. Moonshot says the weights are due by July 27.
Why does K3’s parameter count need qualification?
K3 uses a sparse mixture-of-experts architecture, so only part of its 2.8 trillion parameters is used for each token. Without the active parameter count, its computing needs cannot be inferred from the total alone.
Source: Thorsten Meyer AI
