TL;DR
Moonshot AI released Kimi K3 on July 16 with 2.8 trillion total parameters and API prices matching Claude Sonnet 5’s standard rates. Independent testing places it near leading Western models, but promised weights, licensing terms and a technical report remain pending.
Moonshot AI released Kimi K3 on July 16, pricing the new artificial intelligence model at $3 per million input tokens and $15 per million output tokens—the same standard rates cited for Anthropic’s Claude Sonnet 5. The launch matters because it positions a leading Chinese model as a capability competitor at Western prices, rather than primarily as a cheaper alternative.
Moonshot describes K3 as its most capable model to date, with 2.8 trillion total parameters in a sparse mixture-of-experts architecture. The company says the model routes 16 of 896 experts per token, supports text, image and video input, and offers a maximum context window of 1,048,576 tokens.
K3 is available through the Kimi app, Playground and API. Moonshot’s published API rates are $3 for one million input tokens, $15 for one million output tokens and $0.30 for cached input. Thorsten Meyer AI estimates that the input and output rates are about five times those of the K2 family. Anthropic’s temporary Sonnet 5 introductory rates of $2 and $10, reportedly running through August 31, would make K3 50% more expensive during that period.
Artificial Analysis’ Intelligence Index version 4.1 gave K3 a reported score of 57.1, 2.8 points behind Claude Fable 5 and 1.8 points behind GPT-5.6 Sol Max in the cited configuration. Its long-horizon score reportedly rose by 732 Elo points from K2.6, reaching 1,547, while K3 placed first on Design Arena. These are independent test results reported shortly after release, while Moonshot’s separate benchmark figures remain self-reported.
Western Pricing Resets Competition
The price is a direct test of whether customers will treat Chinese frontier models as peers, rather than discounted substitutes. For the past two years, Chinese developers often attracted users through lower API costs and downloadable weights. Moonshot is now asking buyers to compare K3 with Western systems on performance, reliability and deployment options.
The launch also adds pressure on Western model providers. If K3’s reported performance holds across wider testing, vendors may face a rival offering near-frontier capability with promised downloadable weights at the same API price. That could affect enterprise purchasing decisions, model-routing services and national strategies built on the assumption that Chinese laboratories remain behind at the highest performance tier.
As an affiliate, we earn on qualifying purchases.
K3 Arrives Ahead of Forecasts
Thorsten Meyer AI said models at K3’s reported performance level had been expected in early 2027, putting the release about six months ahead of that forecast. The assessment depends on early benchmark results, and no broad consensus about that timeline has yet been established.
K3 is also much larger by total parameter count than Moonshot’s roughly one-trillion-parameter K2 family and the cited 1.6-trillion-parameter DeepSeek V4-Pro. Its scale complicates claims that Chinese laboratories advanced mainly through doing more with less computing power. Because K3 uses sparse expert routing and Moonshot has not disclosed the active parameter count, however, total parameters do not reveal its computing cost.
“Our most capable model to date, with 2.8 trillion parameters.”
— Moonshot AI, in launch material cited by Thorsten Meyer AI
Licence and Compute Details Pending
K3’s promised weights were not available at launch, so descriptions of it as an open or open-source model remain premature. Moonshot says the weights will arrive by July 27, but the licence has not been published. Until its terms appear, users cannot determine whether commercial use, modification and redistribution will be permitted without material restrictions.
Moonshot has also not released a full technical report or disclosed the number of active parameters. The million-token context figure is a stated maximum, while some access tiers may have lower limits. Only the Max reasoning setting was available at launch, leaving the performance and pricing effects of other reasoning levels unknown.
July 27 Release Faces Scrutiny
Attention will shift to Moonshot’s promised July 27 weight release. Developers will examine the licence, hardware needs, model files and deployment requirements to determine whether K3 can be used outside Moonshot’s hosted services and whether its 2.8-trillion-parameter scale is practical for independent operators.
Further independent testing will show whether K3’s early scores hold across coding, reasoning, multimodal work and long-context tasks. Buyers will also watch real-world latency and reliability, while competitors may respond through price changes, stronger open-weight releases or revised product tiers.
Key Questions
When did Moonshot AI release Kimi K3?
Kimi K3 was released on July 16, 2026, through the Kimi app, Playground and API. Its downloadable weights are promised for July 27.
How much does Kimi K3 cost?
Moonshot lists K3 at $3 per million input tokens, $15 per million output tokens and $0.30 per million cached input tokens.
Is Kimi K3 open source?
Not at this stage. The weights and licence were unpublished at launch, so the extent of permitted use cannot be confirmed until Moonshot releases the licensing terms.
How does K3 compare with leading Western models?
Early Artificial Analysis results place K3 2.8 points from the cited leader and near GPT-5.6 Sol Max. Those results support a near-frontier performance claim, but wider testing is still needed.
Why is K3’s pricing drawing attention?
K3 uses the same $3 input and $15 output rates cited for Claude Sonnet 5’s standard pricing. That choice suggests Moonshot expects customers to judge K3 on capability rather than a large discount.
Source: Thorsten Meyer AI
