We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hacker News·22d ago·5 min read
The real prices of frontier models. Tokens * Price, right?
TL;DR
Comparing AI model pricing based solely on dollars per million tokens is misleading due to varying tokenization methods. For instance, the same TypeScript file results in 1.73 times more tokens on Claude than on GPT-5.x, leading to higher costs.
✦ Why It Matters
Evaluate the tokenization method of your chosen AI model to accurately estimate project costs.
Key Takeaways
- Always check the token count for your specific content on different models.
- Consider the impact of tokenizer changes on your project costs.
- Use tools to measure tokenization differences before selecting a model.
- Budget for potential hidden costs associated with tokenization discrepancies.
Full Summary
AI model pricing often presents a single cost per million tokens, but this figure can be deceptive due to differences in how models tokenize text. Tokenization is the process of breaking down text into smaller units, or tokens, which can vary significantly between models.
For example, a TypeScript file is tokenized into 681 tokens by GPT-5.x but into 1,178 tokens by Claude's latest tokenizer, resulting in a 73% higher cost for the same content. Additionally, Claude's newer tokenizer increases token counts by about 31% compared to its previous version, despite maintaining the same advertised price.
This hidden cost can lead to unexpected expenses for developers using AI coding agents. Understanding these nuances is crucial for accurate budgeting and cost management in AI projects.