Ask three questions in this order and most of the model list falls away: what does a wrong answer cost, what does one task cost, and what may leave your network.
1. What does a wrong answer cost?
If a test catches the mistake, the cheap model is fine. If a person catches it, you are paying that person, and the model's price is the smaller number.
Anthropic's cheapest current model, Haiku 5.5, costs $0.10 per million input tokens and $0.50 per million output. Opus 5.5 costs $4 and $20. That is forty times more, per token, for the same API call.
2. What does one task cost?
Per token is the wrong unit. Per task is the right one: input tokens plus output tokens, times the retries.
A reasoning model writes far more output than it reads. At $50 per million output tokens, a model that thinks for 3,000 tokens before answering spends more than a cheap model that answers in 200 and gets replaced once.
| Model | Input / MTok | Output / MTok | 1,000 tasks at 2k in, 500 out |
|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.45 |
| Claude Sonnet 5.5 | $2 | $10 | $9 |
| Claude Opus 5.5 | $4 | $20 | $18 |
| GPT-6 Astra | $10 | $50 | $45 |
3. What may leave your network?
Retention and residency are decided per provider, not per model, and they are not the same everywhere. Anthropic's zero data retention excludes its Fable and Mythos models, which require 30-day retention. OpenAI's applies to some endpoints and not others. If your answer to this question is "nothing", the model list is now one or two entries long.
The order matters
Most teams ask question two first, because price is the number on the page. Ask question one first: it decides whether the price matters at all.