Token Pricing

Estimating a bill is arithmetic. Suppose a model charges $2 per million input tokens and $10 per million output tokens (illustrative numbers, not a quote). A support reply that reads 3,000 tokens of instructions and context and writes 500 tokens costs 3,000 × $2/1M plus 500 × $10/1M, or $0.011. At 100,000 replies a month that is about $1,100. Real prices change often and differ by model, so check the provider’s current page before budgeting.

Output usually costs more than input because generating each token is a separate, sequential step (see prefill and decode). Reasoning models can also bill their thinking as output tokens, so a short visible answer may be expensive. Providers discount repeated input through prompt caching and non-urgent work through batch inference.

The trap is reading the per-token rate as the cost of the job. An agent that calls a model 40 times per task, rereading its growing context on each call, can cost more on a cheap model than a single call on an expensive one. In episode 67 at 11:38, Cisco’s Jeetu Patel argues that price per token may become a vanity metric for exactly that reason: long-running agents use more tokens even as the price falls, so the number that has to go down is the cost per unit of outcome. Episode 67’s Svix read is at 24:06.

Measure cost per task alongside token prices, and watch for token maxxing, where token volume is rewarded for its own sake. See token for what is being counted.