How to Count Gemini Tokens Locally

2026-07-01 · Source: HackerNoon · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, extended

Summary

This article details how to count Gemini model tokens both locally and through the API, using the "google-genai" Python SDK version 2.9.0 or newer. It introduces the LocalTokenizer class for offline text token estimation, demonstrating its use with the "gemini-3.1-flash-lite" model. The content explains multimodal tokenization math: images consume up to 280, 560, 1120, or 2240 tokens per image based on resolution; audio uses 25 tokens per second; video combines audio tokens with frame tokens (default 1 FPS, 70 tokens/frame for low/medium resolution, 280 for high); and PDFs use 280, 560, or 1120 tokens per page. Benefits of local token counting include offline operation, reduced API calls, lower latency, cost control, and enhanced privacy. For precise billing, the article advises relying on "usage_metadata" from API responses.

Key takeaway

For AI Engineers managing Gemini model deployments, integrating local token counting is crucial for optimizing costs and performance. You should implement the LocalTokenizer from the "google-genai" SDK to pre-validate prompt sizes offline, reducing API latency and avoiding rate limits. This approach enables more precise budget forecasting and allows for dynamic routing of requests based on input token volume, especially for multimodal content, ensuring efficient resource utilization and data privacy.

Key insights

Local tokenization for Gemini models offers offline estimation, cost control, and privacy benefits for multimodal inputs.

Principles

• Tokenization is an information compression codec for LLMs.
• Multimodal inputs are processed by specialized tokenizers with distinct calculation rules.
• API "usage_metadata" is the single source of truth for billing.

Method

Install "google-genai[local-tokenizer]>=2.9.0", initialize LocalTokenizer with "model_name", then call "tokenizer.count_tokens(contents)" for local estimation.

In practice

• Use LocalTokenizer to audit sensitive data token counts offline.
• Route requests to different models based on local token count estimates.

Topics

• Gemini API
• Tokenization
• Multimodal AI
• Local Tokenizer
• Cost Optimization
• Google Gen AI SDK

Best for: AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Prepay for the Gemini API to get more control over your spend — Prepay billing for Gemini API enhances spend predictability and simplifies management for developers.
• You can't trust one token number across your tools. Here's the guide to a dashboard that keeps Codex, Claude, and ChatGPT honest. — A token burn dashboard reveals AI usage habits, improving outcomes and expanding imaginative computing beyond mere cost tracking.
• How I Stopped Being an “Old-school Dev” to Create an Artificial Intelligence Monitor with Gemini — Embracing AI-assisted development necessitates proactive token consumption monitoring to manage LLM API costs effectively.
• We’re expanding Gemini in Chrome to users in Asia Pacific. — Gemini in Chrome integrates AI for personalized browsing, content summarization, and cross-app functionality.
• use this to get the most out of free ai api — Distribute AI API requests across multiple providers using round-robin to prevent rate limits and ensure workflow continuity.
• How to Get 10x Better Results with Google Gemini — Customizing Google Gemini through settings, personalized instructions, and Gems significantly enhances its utility and relevance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by HackerNoon.