Blog

Several providers charge a flat rate up to a threshold and a multiplied one past it — applied to the whole request, not the overage. Crossing that line by a single token roughly doubles the input cost of everything you sent.

A system prompt and a set of tool definitions get billed on every single call in a multi-turn or agentic session, not once — and JSON schemas tokenize worse than prose. That fixed cost is often larger than the message a user actually typed.

GPT-4 Turbo and GPT-3.5 Turbo both leave OpenAI's lineup on 23 October 2026. The named replacements aren't drop-in on price or tokenizer — one gets cheaper and vastly roomier, the other gets pricier for a window many workloads won't use.

The two major providers price caching on opposite models. OpenAI's has no write premium, so any cache hit saves money. Anthropic's charges to write and refunds on read, which means a break-even hit rate — and a low-traffic endpoint that pays more than it saves.

A chat feature bills for its entire history on every turn, so cost grows with the square of the conversation length rather than in step with it. Here is the arithmetic, and the three ways to bend the curve.

Tokens are not words, characters, or syllables — they are the output of a compression algorithm that was fit to a training corpus. Understanding how that algorithm works explains almost every surprising count you will ever see.

The cheapest per-token model is not reliably the cheapest bill, and the most capable one is not reliably worth it. A short decision procedure based on the shape of your workload rather than on benchmark scores.

A context window is a hard ceiling on input and output combined, and it varies by a factor of sixty across current models. Here is what each one actually holds, and why filling it is rarely the goal.

A forecast you can defend takes four numbers and about ten minutes. The estimates that go badly wrong usually miss the same three things — and none of them appear on a pricing page.

The gap between the cheapest and most expensive model on an identical request is roughly 58x. Here is the arithmetic on a realistic workload, and why the cheapest per-token price is not always the cheapest bill.

The same amount of information can cost two or three times as much when it arrives as code instead of English. The reason is in how tokenizers were built, and most of the overhead is avoidable.

Most prompts carry 20–40% padding that contributes nothing to the answer. These are the edits that shrink the bill while leaving the output intact.

Every provider ships its own tokenizer, so the same paragraph can cost meaningfully more on one model than another. Here's what actually drives the number.