Per-token price does not equal per-answer price: Gemini and Mistral have cheap per-token prices, but with an average of as many as 38 steps per question, their per-answer cost is comparable to or even higher than the currently used model.