Claude Sonnet 5は、トークン単価が安いにもかかわらずOpus 4.8より高くつく

claude sonnet 5 scored 53 points on the artificial analysis intelligence index, tying gpt-5.5 (high) and beating opus 4.8 on some agent-based tasks, but it consumes significantly more tokens to achieve these results. an average task costs $2.29 with sonnet 5 versus $1.97 with opus 4.8 because the model runs three times as many agent loops and outputs 40% more tokens than sonnet 4.6. this follows anthropic's pattern of hiding price increases behind new tokenizers or agentic designs while competitors like deepseek v4 pro offer lower prices. developers now face higher effective costs despite static official rates, complicating model selection.