According to OpenAI's official data, Luna's input and output prices have decreased from $5 and $30 per million tokens to $1 and $6 respectively, a reduction of 80%. Claude Opus 5's input and output prices are $5 and $25 per million tokens, respectively, half the price of Fable 5, yet on par with its predecessor, Opus 4.8. This represents "quality improvement at the same price" rather than a direct price reduction. Furthermore, DeepSeek will actually raise the price of its V4 model on August 17th, indicating that open-source models are not limited to continuous price reductions.
Market focus should not be limited to the decrease in token prices. It can extend to the question of whether, as model capabilities converge, customers can continuously switch vendors, making it more difficult for model companies to maintain their pricing power and trillion-dollar valuations. However, this may benefit the application layer; for Nvidia and data centers, it depends on whether the growth in usage can outpace the decrease in unit price. Therefore, it is more worthwhile to observe total token expenditure, model gross margin, and GPU utilization: if the token growth brought about by proxy AI exceeds the price reduction, the demand for computing power will still expand; otherwise, the pressure will be transmitted along the model vendor, cloud vendor, and GPU financing chain.
