The cost of a reasoning model isn't measured in tokens. It's measured in the wall-clock seconds your GPU is pinned to a single request. A standard model predicts in roughly one second; a reasoning model can hold the same GPU for thirty seconds while it works through interleaved thinking, tool calls, and self-correction. That ratio is the actual bill — your concurrent-user capacity drops 30×, your P95 latency goes nondeterministic, and the tokens-per-million number on your invoice is the symptom, not the disease.
Inference scaling means cost stops being linear with input size. The TDS piece walks through where it shows up: chain-of-thought decomposition burning thousands of tokens on simple tasks (the classic "burn tokens to add 1 to 9900" reasoning loop), GPU memory occupancy stretching from sub-second to 30s+, and P95 latency variance that "makes applications feel broken" through timeouts. A concrete case study from the article: shifting simple work off a reasoning model saved $2,030/day — $3,000 down to $970, a 68% cut — without affecting task quality. The lesson is that your reasoning model is not the cheap one for everything; it's the expensive one that's worth it sometimes.
This is why every frontier provider is now selling routing as a product. Claude Sonnet 4.5 + Haiku 4.5, OpenAI o3 + gpt-4.1, Gemini 2.5 Pro + Flash — the routing tier exists because the cost shape of reasoning vs. non-reasoning is genuinely different, and trying to hide that from builders just produces nasty bills. The interesting reframe in the piece: stop measuring "dollars per million tokens" and start measuring "cost per successful task." A reasoning model that solves a problem in 40K tokens but eats two retries is more expensive than a smaller model that nails it in 2K. Your invoice doesn't show this; your task-completion ratio does.
Three things you can do this week. First: classify your traffic into Use / Maybe / Avoid for reasoning — math, planning, multi-step debugging are Use; extraction, formatting, simple lookups are Avoid. Second: set hard caps on reasoning tokens, retries, and total request time so a thinking-trap loop doesn't eat your budget overnight. Third: log per-request `tokens × wall-clock-seconds × success-bool` and look at the cost-per-successful-task distribution, not the average token cost. The reasoning model is a real tool — it just isn't the right tool seventy percent of the time you'll be tempted to reach for it.