What it actually costs

The number that should change your mind

Input prices across the models on this page span about 8,824 times. That is not a small difference in quality for a small difference in price — it is a difference of three orders of magnitude, for models that all answer the same kind of question.

Output costs several times more than input on nearly every model. If you are generating long answers rather than reading long documents, your bill is dominated by output tokens. Ask for less, and pay less.

What actually drives your bill

• Output length, not input length, on most tasks. A terse instruction that produces a 200-token answer costs a fraction of a chatty one.
• Re-sending context. In a long conversation, the whole history is charged again on every turn. Long chats are quadratically expensive, which is why harnesses manage context carefully.
• Choosing a frontier model for a routine job. The cheapest paid models here are perfectly adequate for classification, extraction and formatting — tasks that make up most of the volume in real systems.

How to buy credits

Every route differs, and this is where people waste money:

• Directly from a lab — an account and a card per provider. Best price, worst convenience, one bill per vendor.
• An aggregator — one account and one credit balance reaching hundreds of models. Slightly more expensive per token, and worth it while you are deciding.
• Cloud platforms — the model is one line item on an existing cloud bill, which suits companies that already have one.
• Free tiers — real, and usually paid for with your data. Never send anything confidential through one.
• Prepaid credits — most providers let you buy credit rather than attach a card. Do that first: it caps the damage from a runaway loop.