OpenRouter finds GPT-5.5 costs surge

This is the kind of pricing shift that changes default model choices fast: GPT-5.5 may be smarter, but for many teams it is materially more expensive in practice, not just on paper.

• –OpenRouter’s switcher-cohort method is useful because it compares the same users before and after the model swap, which is closer to production reality than synthetic benchmarks
• –The efficiency gain is workload-dependent: prompts above 10K tokens saw 19-34% fewer completion tokens, while shorter prompts often got longer outputs instead
• –For short and medium chats, the cost penalty is harsh enough to make GPT-5.5 a bad default unless the quality lift is obvious
• –For long-context workflows, the reduced verbosity softens the blow, but it does not erase the price increase
• –The takeaway for developers is simple: route GPT-5.5 selectively, not blindly, and measure cost by prompt bucket instead of averaging everything together