Price per 1M tokens is meaningless
janilowski.pl
July 6, 2026
4 min read
🔥🔥🔥🔥🔥
54/100
Summary
Comparing AI models by price per 1M tokens can be misleading due to differences in tokenization methods among various labs. Each tokenizer affects how text is divided into tokens, which can significantly influence overall costs.
Key Takeaways
- Comparing AI models solely by price per 1M tokens is misleading due to differences in tokenization methods across models and labs.
- Token efficiency varies significantly among AI models, impacting the overall cost of usage despite lower nominal prices per token.
- The cost per benchmark task can differ greatly, with some models being more cost-efficient than others despite higher token prices.
- DeepSeek V4 Pro demonstrates high cost-efficiency, achieving low costs per task despite lower performance on intelligence benchmarks.
Coverage & discussions
Hacker News · janilowski.pl
Community discussions
Hacker News discussion ↗
Historical count: 5 · 5 saved in sample · Historical sample
Analysis ready
Community Sentiment
Positives
- Some commenters recognize that cost per benchmark task provides valuable insight, emphasizing that knowing when a model is ineffective saves time and resources.
- The nuanced discussion about token generation versus prompt processing speed highlights the importance of efficiency in local LLMs, driving home the need for smarter models that meet user goals quickly.
Concerns
- Many feel the current focus on cost per token is misleading, arguing it fails to account for the complexities of model performance and real-world application.
- The skepticism around black box models is palpable, with users lamenting the unpredictable nature of token consumption, leading to confusion and potential waste.
