TL;DR
Moonshot AI released Kimi K3 on July 16, pricing it at $3 per million input tokens and $15 per million output tokens, matching Claude Sonnet 5’s list price. Independent testing places K3 close to leading models, but its weights, licence, technical report and active parameter count are not yet public.
Moonshot AI released Kimi K3 on July 16, offering its new 2.8-trillion-parameter model through the Kimi app, Playground and API at $3 per million input tokens and $15 per million output tokens. Independent testing places K3 near leading Western systems, while its price matches Claude Sonnet 5’s standard list rate, shifting the competitive focus from low-cost access toward capability at comparable prices.
Kimi K3 scored 57.1 on the Artificial Analysis Intelligence Index v4.1, according to figures cited by Thorsten Meyer AI. That placed it 2.8 points behind Claude Fable 5, listed at 59.9, and 1.8 points behind GPT-5.6 Sol Max at 58.9. The tracker also recorded a 732-point Elo increase over Kimi K2.6, bringing K3 to 1,547, while Design Arena ranked it first. These results are independent of Moonshot, but they represent an early evaluation conducted about one day after release.
Moonshot describes K3 as a sparse mixture-of-experts model with 2.8 trillion total parameters. Its architecture routes 16 of 896 experts for each token and uses Kimi Delta Attention and Attention Residuals. The company lists a maximum context window of 1,048,576 tokens and support for text, image and video input. The active parameter count, which would offer a clearer view of the computation used for each token, has not been disclosed.
Pricing marks an equally material change. K3 costs about five times as much as the preceding K2 family, based on the approximate prices cited in the source material, and is described by Thorsten Meyer AI as the most expensive Chinese model released to date. K3 currently costs more than Sonnet 5 during Anthropic’s temporary introductory rate of $2 for input and $10 for output, which is scheduled to expire on August 31.
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.
China’s Model Discount Is Fading
The release challenges the established commercial pitch for Chinese models: capable systems offered at a steep discount or through downloadable weights. By setting K3 at the same list price as Claude Sonnet 5, Moonshot is signaling that it expects buyers to compare the systems on quality, reliability and deployment options, rather than treating K3 as a cheaper substitute.
The timing also matters. Thorsten Meyer AI characterized K3 as reaching this performance tier about six months earlier than expected, based on forecasts that placed comparable Chinese systems in early 2027. That timing is an analytical claim rather than a documented industry consensus. Still, K3’s early independent scores suggest that the measured frontier gap has narrowed, increasing pressure on Western providers competing in coding, reasoning and multimodal workloads.
K3’s scale complicates claims that hardware restrictions have forced Chinese laboratories to rely mainly on smaller, more efficient designs. A 2.8-trillion-parameter sparse model is not directly comparable with a dense model because only part of the network is used for each token. Without the active parameter count or technical report, the release cannot establish how much training or inference computation K3 requires, or what it indicates about the effect of export controls.
AI language model API access
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From K2 Pricing to Frontier Ambition
Moonshot’s K2 models were associated with prices near $0.60 per million input tokens and $3 per million output tokens. That gap helped support the broader view that Chinese laboratories would compete through lower prices while Western companies retained an advantage at the high end. K3 replaces that positioning with premium API pricing and frontier-level claims.
Moonshot announced K3 as its most capable model to date and said it contains 2.8 trillion parameters, correcting reports that described the total as 2.7 trillion. The company has also presented it as an open-weight release, but the weights were not available at launch. Unlike models released immediately under documented permissive terms, K3 cannot yet be independently inspected, hosted or evaluated under its final licence.
“Our most capable model to date, with 2.8 trillion parameters.”
— Moonshot AI launch materials
large language model development kit
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
K3’s Open-Weight Terms Remain Missing
Several parts of K3’s release remain unverified. Moonshot says the model weights will arrive by July 27, but it has not published the licence governing their use. The technical report and active parameter count are also unavailable, leaving researchers unable to check training methods, computational requirements or reuse restrictions.
The advertised one-million-token context is a maximum capability rather than guaranteed access across every service tier; the Moderato configuration is reportedly capped at 256,000 tokens. Only the Max reasoning setting was available at launch. It is also unclear how K3’s benchmark performance will hold up under broader independent testing, production traffic and comparisons using matched reasoning and cost settings.
AI model evaluation tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
July 27 Will Test Moonshot’s Claims
Attention now turns to Moonshot’s promised July 27 weight release. Researchers and developers will examine the licence, test whether the downloadable model reproduces hosted results and measure its hardware demands. Publication of the technical report and active parameter count would help establish K3’s real efficiency and deployment costs.
Further independent evaluations will also show whether the initial 57.1 index score is stable. Buyers will then be able to compare K3 with Claude Sonnet 5 and other models using sustained performance, latency and total operating cost, rather than launch-day benchmarks alone. Until those results and terms are available, K3’s competitive position remains promising but provisional.
AI model benchmark testing software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Kimi K3?
Kimi K3 is Moonshot AI’s newest multimodal reasoning model, released through the Kimi app, Playground and API. Moonshot says it has 2.8 trillion total parameters and accepts text, images and video.
How much does Kimi K3 cost?
The API price is $3 per million input tokens, $15 per million output tokens and a reported $0.30 rate for cached input. Its standard input and output pricing matches Claude Sonnet 5’s list rate.
Is Kimi K3 already open-weight?
No. Moonshot has announced an open-weight release by July 27, but the weights and licence were unavailable at launch. Until both are published, outside users cannot verify the usage terms or independently host the promised release.
How close is K3 to leading AI models?
Artificial Analysis gave K3 a score of 57.1, 2.8 points behind the highest result cited in the source material. The figure indicates a narrow gap in that test suite, but broader evaluations are still needed.
Why is the pricing change important?
K3 is about five times more expensive than its predecessor and priced alongside a Western mid-tier competitor. That suggests Moonshot wants K3 judged as a direct capability rival, rather than mainly as a discounted alternative.
Source: Thorsten Meyer AI
