AI model token counts vary by 20% for identical requests

A developer observed a significant discrepancy in token counts for identical API requests routed to different AI models. One request, processed by a mid-tier model, resulted in 246,525 input tokens, while a nearly identical request (0.3% larger in byte size) sent to a smaller model was billed for only 196,892 tokens, a 20.1% reduction. This suggests that token counts are model-scoped, not solely a property of the input bytes, implying that direct cost comparisons based on "dollars per token" across different models can be misleading. AI

IMPACT Highlights that direct cost comparisons between AI models can be misleading due to model-specific tokenization.

RANK_REASON Developer's observation and analysis of AI model behavior, not a formal release or research paper.

Read on dev.to — LLM tag →

• coding-agent CLI
• mid-tier model
• small model

• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →