Claude Sonnet 5 ranks fifth on the Artificial Analysis Intelligence Index, outperforming the more expensive Claude Opus 4.8 on agent-based tasks. However, the model consumes roughly 40 percent more tokens per task than Sonnet 4, effectively doubling real costs while Anthropic maintains unchanged list prices.
This represents a continuation of Anthropic's pricing strategy. By keeping per-token rates static while models become less efficient, the company obscures price increases from customers. Users see no change in published rates but pay substantially more for identical work.
Token efficiency matters because API pricing operates on consumption. A model that processes the same task with fewer tokens costs less, regardless of headline pricing. Anthropic's Sonnet 5 breaks this pattern. Developers switching from Sonnet 4 will face doubled expenses without any price change announcement.
The pattern reveals a pricing tactic that differs from direct rate increases but achieves similar revenue growth. Anthropic avoids the negative publicity of raising announced prices while capturing increased value from improved models. Customers comparing list prices see no difference. Only after deployment do they discover actual costs have risen.
This approach creates friction in the market. Developers budgeting API costs based on published rates will encounter unexpected expenses. Competitors like OpenAI typically pair model improvements with either stable pricing or transparent rate changes. Anthropic's strategy generates the appearance of flat pricing while actual per-task costs climb.
The efficiency decline raises questions about optimization priorities. New models could theoretically achieve better performance with stable or improved token efficiency. That Sonnet 5 trades efficiency for capability suggests Anthropic optimized for benchmark performance rather than cost-effectiveness for customers.
For enterprises standardizing on Claude, this creates planning uncertainty. Each model release potentially doubles costs without warning. The pattern incentivizes careful testing before upgrading, as the better