How to read official API prices: start with per-million-token rates, then cache and tiers

The scene

You want to compare providers or estimate a task’s cost without relying on a third-party roundup.

Tools involved

What they did

1. Use the provider’s own pricing page, not a remembered older rate. The figures on this page were checked on 10 Nov 2026.
2. Copy three numbers first: input, output, and cache-hit prices, usually per million tokens (Anthropic writes MTok). Output is often several times the input rate, so a short draft-then-edit pass costs far less than repeatedly sending a long document through a strong model.
3. A spread of official standard rates, all per million tokens: DeepSeek flash off-peak cache-miss input $0.15 and output $0.60; Gemini 3.5 Flash-Lite $0.30 / $2.50; OpenAI gpt-5-mini $0.25 / $2.00 and gpt-5.4-mini $0.75 / $4.50; Claude Sonnet 5.5 $2 / $10; xAI’s site lists grok-4.7 at $2.00 / $6.00; Claude Opus 5.5 $4 / $20; OpenAI short-context gpt-5.4 $2.50 / $15.00.
4. Then check discounted tiers. OpenAI lists Batch separately; gpt-5.4-mini Batch is $0.375 / $2.25, about half the listed standard rate. Anthropic’s consumer pricing page also states you can save 50% with batch processing. Cache matters too: Sonnet 5.5 cache hits are $0.10 per million tokens, well below the $2 base input price.
5. Estimate with expected input tokens times the input rate plus expected output tokens times the output rate, and include the system prompt, any long material you paste repeatedly, and thinking tokens. That is a ceiling sense, not a billing promise.

Copyable prompt

I will paste official pricing lines I copied today and a rough description of my weekly workload (typical prompt length, how often I re-send the same context, and whether I can wait for batch). Help me compare the options using only those pasted figures. Show the arithmetic in a table. If a number is missing from my notes, say so — do not fill it in from memory.

Rewritten here for you to try; not a verbatim excerpt from elsewhere. What it asks：Ask the model to compare options using only the official figures you pasted today and your own usage notes, and to say so when a number is missing.

Things to watch

Third-party roundups, screenshots, and older posts go stale. The same model may also have a long-context surcharge (OpenAI lists separate long-context prices for some models), regional rates, tool-call fees, and per-request search or fetch charges (DeepSeek lists Search and Fetch at $0.75 per 1,000 successful requests). A free tier marked “free of charge” is not unlimited; Google describes “generous limits,” and the cap is whatever your account and the current page say. This is not purchasing advice.