Peter Liu Claims Claude Opus Uses More Tokens Than GPT
Co-founder of Compound AI posts comparison of token usage for English and code texts.
Peter J. Liu, co-founder of Compound AI and former Research Scientist at Google Brain and DeepMind, tweeted that Claude models require significantly more tokens than GPT models to represent identical text. He reported that Opus 5 uses 50 percent more tokens for English samples including Pride and Prejudice and Moby Dick, and 70 percent more tokens for a 100KB TypeScript file in VSCode. Liu compared the token counts between GPT 5.6 Sol and Opus 5 when discussing per-token costs.
Combined views
821
1 post, first seen 4h ago
Peter Liu Claims Claude Opus Uses More Tokens Than GPT
Co-founder of Compound AI posts comparison of token usage for English and code texts.
Peter J. Liu, co-founder of Compound AI and former Research Scientist at Google Brain and DeepMind, tweeted that Claude models require significantly more tokens than GPT models to represent identical text. He reported that Opus 5 uses 50 percent more tokens for English samples including Pride and Prejudice and Moby Dick, and 70 percent more tokens for a 100KB TypeScript file in VSCode. Liu compared the token counts between GPT 5.6 Sol and Opus 5 when discussing per-token costs.
Reactions
1 postPJ 
Peter J. Liu@peterjliuClaude models cost a lot more per token than GPT if you adjust for the fact that you need significantly more tokens to tokenize the *same* text. Opus 5 uses 50% more tokens to tokenize English (Pride and Prejudice + Moby Dick), and 70% more for VSCode ( 100KB Typescript).…
- likes: 8
- replies: 4
- bookmarks: 1
- reposts: 1