Claude Sonnet 5 Arrived Twice as Expensive Because It Can't Stop Thinking

Independent testing platform Artificial Analysis ran the brand-new Claude Sonnet 5 through their Intelligence Index v4.1, revealing that the model burns roughly twice the budget of its predecessor on maximum reasoning effort. Testing Sonnet 4.6 on max effort cost $1.14 per task, while Sonnet 5 racked up a hefty $2.29 bill for a modest bump from 47 to 53 points on the overall benchmark scale.

This financial bloat isn't caused by a subscription hike. Anthropic kept baseline API pricing rock-solid at $3 per million input tokens and $15 per million output tokens. The real plot twist is that the model simply became dramatically more verbose: it spits out 40% more output tokens and executes triple the reasoning steps on agentic tests like GDPval-AA and AA-Briefcase, consuming an average of 69,000 tokens per single prompt like a stressed intern trying to justify overtime.

Because of this chatterbox behavior, Sonnet 5 accidentally made its flagship sibling look like an unbelievable bargain. The heavier Opus 4.8 scores 56 points on the benchmark while billing just $1.80 per task—meaning the supposedly economical middle-tier Sonnet is currently about 15% more expensive to run than the premier flagship model.

The extra steps do deliver value in specific niches, though. In pure enterprise workflow tasks like document analysis and office operations, Sonnet 5 outpaces Opus 4.8 and only trails the elite Fable 5. However, in abstract scientific benchmarks like CritPt, the bloated token usage only pushed it to a 17% accuracy score, leaving it trailing behind GLM-5.2, GPT-5.5, and Opus.

To stop early adopters from fleeing in terror, Anthropic rolled out a temporary promotional discount shaving a third off standard token rates until September 1, dropping costs to $2 input and $10 output. Meanwhile, a new 'xhigh' effort tier mirrors the control knobs of Opus, ensuring that burning enterprise budgets at record speed remains entirely an optional design choice.

The era of measuring AI efficiency by nominal token price is officially bankrupt when models simply invent extra steps to burn more cash. When thinking harder costs double for a tiny bump in raw intelligence, enterprise teams suddenly face the weirdest paradox in software: the luxury flagship just became the budget option.

Source: Artificial Analysis