Count Tokens for a Gemini Request

The Gemini API provides a countTokens endpoint that returns the exact token count for a request — including system instructions, conversation history, and user messages — without consuming your quota or generating a response. This is especially valuable for Gemini given its massive 1M–2M token context windows, where fitting within limits requires precise measurement. This example shows a countTokens request and explains how to interpret the totalTokens response field. Gemini uses the SentencePiece tokenizer, which produces different token counts than GPT's tiktoken or Claude's tokenizer. For multilingual content, SentencePiece typically produces more efficient tokenization for languages beyond English. For code, the counts are similar across tokenizers. Always use Gemini's own countTokens endpoint rather than estimating from another model's token count. For vision-capable requests (Gemini 1.5 Pro and Flash), images are also counted — each image uses approximately 258 tokens for a standard-resolution image. The countTokens endpoint accepts image content in the same format as the generateContent endpoint, so you can get an accurate total count for multimodal requests before submitting them.

Count tokens for this Gemini request:

System instruction: You are a helpful assistant that answers questions about software architecture and design patterns.

User message: Explain the difference between the Repository pattern and the Active Record pattern in software development. When should I use each one? Provide a concrete example in Python for both patterns.

Model: gemini-1.5-flash