The list prices fell 80% in three months — DeepSeek did that

Data scientist Max Woolf tracked every price move from both labs since June: OpenAI cut GPT-5.6 Luna 80 percent in a single July announcement, GPT-6 Sol launched in September at half the price of the model it replaced, and Claude Opus 5.5 landed the same day 40 to 60 percent cheaper than Opus 5. Vercel's gateway data puts the overall cost per token at 43.3 percent of April levels.

The driver, Woolf argues, is not generosity. DeepSeek's Flash V4.1 sells for $0.15 per million input tokens with a cache-read cost a fraction of the incumbents', and 29 inference providers now host it on OpenRouter. With switching costs near zero, the labs are racing the price of intelligence down toward its marginal cost — months before both are rumored to file for IPOs.