Why o3-mini *had* to be free: the coming DeepSeek R1, 2.0 Flash, and Sky-T1 Price War
Unlike Meta executives, we are truly on the fence about whether or not DeepSeek’s “$5.5m model trained with box of scrap GPUs with no MCTS/PRM” is a psyop1. However, we do believe that the price-intelligence Pareto frontier is both closely watched and predictive, as we saw with the Gemini Pro price cut last September, and our updates with new models offer more clarity where the cost of reasoning is headed:
While it was expected that o3-mini due to launch “ in ~a couple of weeks”, a surprise announcement after the R1 release this week (not counting Stargate and Operator) was that o3-mini would also be launched in free ChatGPT:
With Noam Shazeer back at Google actively shipping updates to Gemini 2.0 Flash Thinking - which is available for free with no pricing announced2, and with DeepSeek launching o1-competitive models at 27x cheaper than o1, the price pressure is on.
However, we believe the pressure is EVEN MORE intense for the -mini models than for the full reasoner models. The biggest surprise of the DeepSeek R1 paper wasn’t the performance of R1 itself, which was already preannounced in November.
