Financial news graphic showing Claude Sonnet 5 token rates unchanged despite rising real-world costs, highlighting stable tok

Editorial illustration for Claude Sonnet 5 lists unchanged token rates despite double real cost

Claude Sonnet 5 lists unchanged token rates despite...

Claude Sonnet 5 lists unchanged token rates despite double real cost

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

July 1, 2026 • 2 min read

Claude Sonnet 5 arrived with a lot of fanfare, yet its pricing sheet looks unchanged from Sonnet 4.6: $3 per million input tokens and $15 per million output tokens. The model, however, burns far more tokens to hit the same headline numbers. In an independent test by Artificial Analysis, Sonnet 5 landed fifth on the Intelligence Index, tying with GPT‑5.5 (high) at 53 points, while the top slot, Claude Fable 5, sits at 60.

That six‑point jump over Sonnet 4.6 (47 points) comes at a cost. The same study estimates an average task now costs $2.29 with Sonnet 5, versus $1.97 for Opus 4.8, even though Opus 4.8’s token rates are higher at $5 and $25. At “max” performance, Sonnet 5 uses about 40 percent more output tokens per task than its predecessor, and on agent‑based benchmarks it runs roughly three times as many loops.

The result? A near‑doubling of per‑task expense despite the unchanged headline rates.

Claude Sonnet 5 continues Anthropic's pattern of hiding price increases behind unchanged token rates In an independent test, Claude Sonnet 5 placed fifth and beat the pricier Opus 4.8 on some agent-based tasks.

— Claude Sonnet 5, Claude Sonnet 5 continues Anthropic's pattern of hiding price increases behind unchanged token rates - THE DECODER

Why this matters

We see Claude Sonnet 5 priced the same as its predecessor—$3 per million input tokens, $15 per million output—yet its token usage has roughly doubled, pushing the per‑task cost above that of Anthropic’s earlier flagship. The model slipped into fifth place in an independent benchmark, even outpacing the pricier Opus 4.8 on certain agent‑based tasks, which suggests a performance edge in niche scenarios. However, the spike in consumption means developers may pay more for the same workload, a trade‑off that could curb adoption outside specialized use‑cases.

If the average task now costs $2.29, as Artificial Analysis reports, the headline “same token prices” masks a hidden expense that could affect budgeting decisions for startups and research teams alike. It remains unclear whether the performance gains justify the higher effective price, especially when cheaper alternatives exist.

We should therefore scrutinize token efficiency as a core metric, not just raw scores, before integrating Sonnet 5 into production pipelines. The pattern of unchanged rate tables coupled with increased consumption invites caution; transparency around true cost structures will be essential for informed development choices.