CJ Wright

Sam Altman shared that OpenAI is losing money on its o1 pro model ($200/month) because of usage, but... because of how reasoning models work, $200/month is likely inherently too low of a price point. Based on known stats about reasoning models, it should cost ~$1000/month to break even. Here's why. Open-source reasoning models show that the "reasoning" part of the model produces about 10,000 hidden, intermediary "reasoning" tokens (chain of thought, etc.) per 1,000 outputted tokens. In other words, for every 1,000 tokens you see outputted by the model, it's actually generated 11,000 (with 10,000 under the hood). That alone should put the price point for a reasoning model like o1 pro at 11x the standard price ($20) - $220. But then we need to consider the cost of the expanded context length. Typically, models can handle multiple requests (multiple users) at once. However, as the context length of a request gets longer, the amount of concurrent requests the model can handle is reduced - it's estimated that reasoning models reduce the number of possible concurrent users by 75-80%. In other words, only 1/4 - 1/5 as many people can use the model at the same time. Therefore, you need to multiple that cost by 4-5x. Now we're at $880-1100/month.