Gemini 1.5 Pro uses tiered pricing: prompts up to 128K tokens cost $1.25/1M input, but prompts above 128K cost $2.50/1M. If you routinely send 150K-token contexts, you're in the higher pricing tier for the entire prompt — not just the tokens over 128K. Structure your application to stay under 128K when possible.

When you send a PDF to Gemini, it processes each page as an image (~1,290 tokens per page) regardless of whether the page is mostly text. A 50-page PDF costs ~64,500 tokens in visual processing alone. For text-heavy PDFs, extract the text with a PDF parser first and send plain text instead — dramatically cheaper.

Unlike OpenAI's chat format, Gemini's system instruction is a separate field but still counted in your token total for each API call. If you use the same 500-token system instruction on 50,000 calls per day, that's 25M input tokens per day from instructions alone. Enable context caching for static system instructions.

Gemini's Google Search grounding feature fetches live web content and injects it into context. Each search result page adds hundreds to thousands of tokens to your input. Costs scale with how many results are retrieved and their length. Only enable grounding for queries that genuinely require real-time information.

Gemini 2.0 Flash Thinking uses internal reasoning tokens (similar to o1) that are billed at output rates. Complex reasoning tasks can generate 1,000–5,000 thinking tokens before producing the answer. For simple tasks, use Gemini 2.0 Flash (non-thinking) which produces only visible output tokens.

When using Gemini's File API to upload large files (videos, documents) for reuse across calls, the tokens are counted fresh on every API call that references the file — even though the file is stored on Google's servers. File storage is free, but the token cost is paid on each inference.