Input prices run from $0.10 to $10 per million tokens, a hundredfold spread across models that all describe themselves as frontier. Picking the wrong one costs more than any optimisation you will do this quarter.
| Model | Input / MTok | Output / MTok | One call, full window |
|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.10 |
| Claude Sonnet 5.5 | $2 | $10 | $2.00 |
| Gemini 4 Carbon | $2 intro, $4 after | $10 intro, $20 after | $2.00 |
| Claude Opus 5.5 | $4 | $20 | $4.00 |
| GPT-5.5 | $5 | $30 | $5.00 |
| GPT-6 Astra | $10 | $50 | $10.00 |
| Claude Fable 5.1 | $10 | $50 | $10.00 |
The intro price is a trap
Gemini 4 Carbon launched at $2 and $10 and moves to $4 and $20. A budget built on the launch price is wrong by half within a quarter. Put the standard price in the spreadsheet and treat the discount as a temporary gift.
A full window is a bill
At Opus 5.5's $4 per million input tokens, one call that reads a million-token window costs $4. Twenty of those in a session is $80, and the model has not written anything yet.
This is why agent loops are expensive: each step re-sends the conversation. A ten-step task on a 200,000-token history is $8 in input alone.
The cheap tier is not a downgrade
Haiku 5.5 at $0.10 is forty times cheaper than Opus 5.5. For classification, extraction, routing and summarising, it is usually good enough, and the saving funds the calls that need the expensive model.
Route by task. A request that needs the frontier model twice a day should not be paying frontier prices for the other 4,000.