DeepSeek R2 launches — China's open-weights reasoning frontier matches GPT-5 on AIME, undercuts inference cost by 11x
DeepSeek released R2 this week, the long-awaited successor to R1 from January 2025. R2 matches GPT-5 on AIME 2025 (0.94 vs 0.95) and exceeds Claude 4.6 Opus on MATH (0.91 vs 0.89). The model ships under MIT license with full reasoning-trace weights, and DeepSeek's hosted inference prices it 11x below GPT-5's reasoning tier and 7x below Claude 4.6 Opus's extended-thinking tier.
The inference-cost gap is the durable commercial story. DeepSeek's inference economics — running on H800 and Huawei Ascend 910B inference at heavily amortized cost — let the company hold the price point through extended demand. Hosted competitors trying to match R2's evals on Western GPU infrastructure cannot match R2's hosted prices without absorbing material margin loss.
The MIT license is the strategic provocation. R2 ships with no usage restrictions, no commercial-use limit, and no fine-tune restrictions — a stricter open-weights posture than Llama 5.1's modified Community License. The contrast with the Meta open-weights posture sharpens just as Llama 5.1 ships, putting both labs on the open-weights frontier with different governance bets.
DeepSeek — R2 release announcement → · HuggingFace — DeepSeek R2 model card → · MIT Tech Review — DeepSeek R2 sets open-weights reasoning bar →