TL;DR
Moonshot AI released Kimi K3 on July 16 with pricing of $3 per million input tokens and $15 per million output tokens, the same listed rates as Claude Sonnet 5. Independent testing places K3 close to leading models, but its weights, licence, technical report and active parameter count remain unpublished.
Moonshot AI released Kimi K3 on July 16 at $3 per million input tokens and $15 per million output tokens, matching the listed API price of Anthropic’s Claude Sonnet 5. The pricing, about five times that of Moonshot’s previous K2 family, signals that the Chinese developer intends to compete more directly on model performance rather than a steep discount.
Kimi K3 is available through the Kimi app, Playground and API. Moonshot describes it as a 2.8-trillion-parameter sparse mixture-of-experts model that routes 16 of 896 experts per token. It accepts text, image and video inputs and advertises a maximum context window of 1,048,576 tokens.
Independent results cited by Thorsten Meyer AI place K3 at 57.1 on Artificial Analysis Intelligence Index v4.1, compared with 59.9 for Claude Fable 5 using an Opus 4.8 fallback and 58.9 for GPT-5.6 Sol Max. That leaves K3 2.8 points behind the leading tested configuration. Artificial Analysis also recorded a 732-point Elo improvement over K2.6 on its long-horizon tracker, bringing K3 to 1,547.
The price comparison is less favorable during Anthropic’s introductory period. Claude Sonnet 5 is listed at $2 per million input tokens and $10 per million output tokens through August 31, according to the supplied comparison data. During that period, K3 costs 50% more at both ends, despite matching Sonnet 5’s standard listed rates.
K3 Pricing Challenges China Discount
Chinese AI models have often competed through a combination of lower API prices, downloadable weights and competitive benchmark results. K3 changes that calculation: Moonshot is asking customers to pay Western mid-tier rates while presenting the model as a close competitor to leading systems.
Thorsten Meyer AI described the pricing as a larger strategic signal than the benchmark scores, arguing that Moonshot has stopped compensating through discounts. That is an interpretation, not a confirmed company strategy. Still, the published rates create a direct test: customers can compare quality, reliability, latency and deployment flexibility without K3 holding a clear list-price advantage.
AI Systems Performance Engineering: Optimizing Model Training and Inference Workloads with GPUs, CUDA, and PyTorch
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Moonshot Moves Beyond Low-Cost Positioning
Moonshot’s previous K2 family cost about $0.60 per million input tokens and $3 per million output tokens, based on figures supplied by Thorsten Meyer AI. K3 raises those rates roughly fivefold and is described by the publication as the most expensive model released by a Chinese laboratory. That ranking has not been independently verified across every Chinese provider and pricing tier.
K3 also expands Moonshot’s scale from roughly one trillion parameters in the K2 family to 2.8 trillion. The model uses a sparse architecture, so total parameters do not equal the computing load for every token. Moonshot has not disclosed the active parameter count, limiting comparisons with other mixture-of-experts systems.
“Our most capable model to date, with 2.8 trillion parameters.”
— Moonshot AI, in launch material cited by Thorsten Meyer AI
Developing Apps with GPT-4 and ChatGPT: Build Intelligent Chatbots, Content Generators, and More
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Licence and Weights Still Pending
K3 is being described as an open-weight model, but its weights were not available at launch. Moonshot has promised publication by July 27, while the licence and technical report also remain unpublished. Until those materials arrive, claims about commercial reuse, modification rights and independent deployment cannot be verified.
Other open questions include the model’s active parameter count, full inference requirements and performance outside benchmark settings. The one-million-token context figure is a maximum rather than a guarantee across all service tiers; the Moderato configuration is reportedly capped at 256,000 tokens. Only the Max reasoning setting was available at launch, leaving the behavior and cost of other settings untested.
AI model performance benchmarking tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
July 27 Tests Open-Weight Claims
The next milestone is July 27, when Moonshot says it will release K3’s weights. Developers will then be able to examine the licence, hardware demands and independent deployment options. Further third-party testing will show whether K3’s early benchmark position holds across coding, reasoning, multimodal work and long-context tasks, and whether customers accept Sonnet-level pricing.
As an affiliate, we earn on qualifying purchases.
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 already open-weight?
No. Moonshot says the weights will be released by July 27, 2026. The licence was not public at launch, so permitted uses remain unknown.
How does K3 compare with leading models?
Artificial Analysis scored K3 at 57.1 on its Intelligence Index v4.1, placing it 2.8 points behind the leading listed configuration. That is one independent benchmark, not a guarantee of performance in every application.
Why is K3’s pricing drawing attention?
The rates are about five times the K2 family’s approximate prices and match Claude Sonnet 5’s standard list price. This removes much of the price gap commonly associated with Chinese AI models.
Can K3 run on a personal workstation?
The announced model contains 2.8 trillion total parameters, making local operation likely to require substantial hardware even with sparse expert routing. Exact requirements remain unknown because active parameters and deployment documentation have not been published.
Source: Thorsten Meyer AI
