Artificial Analysis, an “Independent analysis of AI models and hosting providers” outlet, have published a chart plotting “Intelligence Index vs. Price” (X post). This places Grok 3 Mini Reasoning at the upper left corner, inside the “Most attractive quadrant”. According to this, it has the best intelligence/price ratio - even at High-Reasoning capacity. The catch: price is measured by token prices, and tokenizer inefficiencies already cast doubt on the chart’s accuracy. Further, the “internal monolog” of reasoning models can vary) - but implicit caching may again alleviate costs in multi-turn chats.

So I repeated some of the English-language visualization tasks from my recent article via OpenRouter:

• Grok 3 Mini does incur more tokens than OpenAI o4-mini even on input, up to 10-17%
• Reasoning may also incur more token use (up to 25%), but may also use less
• output texts are also longer than o4-mini, up to 3x
• (the length of reasoning seems to vary between re-tries, in one case with o4-mini having been 1,5x as large)

Despite the generally higher token use, Grok 3 Mini is much cheaper in total, up to 1/18th of o4-mini. With one unpublished German-language visualization task, Grok 3 Mini incured 1/11th of the cost of o4-mini.