Artificial Analysis' Intelligence Index: Sonnet 5.5 (max) ranks above GPT-6 Astra (max) and behind only Opus 5.5 (max) but has the highest token use of them all

Claude Sonnet 5.5 can achieve top-tier performance, but its cost per task has increased by approximately 50%.

Anthropic's Claude Sonnet 5.5 has achieved the second-highest score on the Artificial Analysis Intelligence Index, narrowly trailing Opus 5.5. While Sonnet 5.5's performance in knowledge work and agentic tasks, particularly Terminal-Bench 4.0, approaches that of top-tier models like Opus 5.5 and GPT-6 Astra, it achieves this with significantly higher token usage. At its maximum effort setting, Sonnet 5.5 consumed approximately 193,000 output tokens per Intelligence Index task, the highest recorded to date, representing a substantial increase over Sonnet 5 and Opus 5.5. Despite this, its pricing remains the same as Sonnet 5. However, when considering cost per task at high effort levels, Sonnet 5.5 is positioned behind Opus 5.5 and narrowly behind GPT-6 Sol. The model also shows weaker performance in factual knowledge and scientific reasoning compared to Opus 5.5.

The performance of Claude Sonnet 5.5, while nearing Opus 5.5, is achieved through an unprecedented level of token consumption. This suggests a trade-off between model efficiency and raw capability at higher effort settings, potentially impacting cost-effectiveness for certain use cases. The comparison with GPT-6 Sol highlights that alternative models may offer comparable or superior intelligence-to-cost ratios, indicating a competitive market where token efficiency is a critical differentiator.

This development signals that while Anthropic is advancing its smaller model class to compete at higher performance tiers, users must carefully consider the operational costs associated with Sonnet 5.5's increased token usage. The AI market continues to emphasize a balance between performance, cost, and efficiency, and Sonnet 5.5's position suggests ongoing innovation focused on maximizing output, even at a higher resource cost, while still being priced competitively at lower effort levels.