The Reaction of Financial Markets to the Releases of DeepSeek R1 and Kimi K3: A Critical Analysis (arXiv (Cornell University))

Summary

• DeepSeek R1 was published on 20 January 2025 as a 671 billion parameter model requiring only 37 billion parameters activated per inference pass.p.5
• DeepSeek R1 was served at API prices of $0.55 per million input tokens and $2.19 per million output tokens, well below OpenAI o1-preview at $15.00 input and $60.00 output.p.5
• Running the Artificial Analysis Intelligence Index evaluation suite cost $2,437.41 with Kimi K3, compared to $5,630.52 for Claude Fable 5 and $3,442.81 for GPT-5.6 Sol.p.7
• During the Kimi K3 reaction window of 15–17 July 2026, share prices of Chinese AI labs Zhipu AI and MiniMax Group fell by 35.1% and 17.1% respectively.p.9
• Total weekly token usage on OpenRouter increased from 0.53 trillion tokens per week in January 2025 to 58.00 trillion in July 2026.p.17

Key insights

1. DeepSeek R1 provided reasoning performance at API prices more than an order of magnitude below contemporary closed frontier competitors.p.6
2. Kimi K3 offered an evaluation cost of $2,437.41, which was 29.2% lower than GPT-5.6 Sol, but 3.8 times higher than Grok 4.5.p.7
3. Market losses following DeepSeek R1 concentrated in compute suppliers with NVIDIA falling 17.0%, whereas losses following Kimi K3 concentrated in rival Chinese AI labs like Zhipu AI.p.9
4. Moonshot AI pursued a Hong Kong IPO at a valuation of approximately $30 billion, supported by $300 million in annual recurring revenue as of June 2026.p.10
5. Closed models accounted for roughly 80% of OpenRouter tokens between May and September 2025 despite costing on average six times as much as open models.p.12
6. NVIDIA quarterly revenue grew from $44 billion in the quarter ending April 2025 to $82 billion in the same quarter of 2026 despite efficiency gains from open models.p.18