The "real cost" of running near-SOTA models is not a secret: you can run local models on your own infrastructure. When you do, you quickly find out that typical agentic coding incurs outsized costs by literal orders of magnitude compared to the simple Q&A chat most people use AI for. All tokens are very much not created equal, and the typical coding token (large model, large noisy context) costs a lot even under best-case caching scenarios.