AI Series #58: Kimi K3 Model Capabilities an Unexpected Surprise for Market

Research evidence excerpt

AI Series #58: Kimi K3 Model Capabilities an Unexpected Surprise for Market

China (PRC) | Internet EquityJulyResearch16, 2026

Exhibit 1 - Input and output prices for flagshipAI Series #58: Kimi K3 Model Capabilities an

models by companies

Models Total parameters Input Price (USD) Output Price (USD)

Kimi-3 2.8T 3.00 15.00Unexpected Surprise for Market DeepSeek-V4-Pro 1.6T 0.87 1.74

Qwen3.7 Max 1.2T 2.50 7.50

GLM-5.2 744B 1.40 4.40

MiniMax M3 428B 0.60 2.40

Hy3 295B 0.06 0.22Moonshot AI released the first open 2.8T model K3 which ranked No. 3 Claude Opus 4.8 Not disclosed 5.00 25.00

in Artificial Analysis and behind Fable5 and GPT-5.6 Sol (Max). It ranked .Source:Claude SonnetCompany,5 NotJefferiesdisclosed 3.00 15.00

Note: Pricing for Qwen is based on official prices after end of

No. 1 in Frontend Code Arena. We view the breakthrough as unexpected promotion, pricing for DeepSeek-V4-Pro is during peak hours

and drives sector innovation amid fast model iterations. According to Token Exhibit 2 - Blended prices for China models vs

Expenditure Index, it reflects continued appeal for a cost-efficient model US models (2Q26)

(USD per million tokens)

16(Exhibit 3), which is positive for China models with strong cost performance 14

(Exhibit 2) 1210

In this report, we provide investors with data points on model pricing by companies, 42

pricing comparison between China and US, model intelligence from Artificial Analysis, token - Input price Output price

consumption in OpenRouter, Token Expenditure Index trends and other industry data for . China US

analysis. Source: Artificial Analysis, Jefferies

Exhibit 3 - Sillicon Data LLM TokenMoonshot AI released the first open 2.8T model and ranked top positions across different Expenditure Index

metrics (Exhibit 5-6).