Claude Sonnet 5: hidden price hikes despite same token rates

Claude Sonnet 5 scored 53 points in Artificial Analysis's Intelligence Index v4.1, tying GPT-5.5 (high) for fifth place, and improving six points over Sonnet 4.6. Despite keeping the same listed token prices as its predecessor, Sonnet 5 consumes materially more tokens, pushing average task cost above competitors.

How did Claude Sonnet 5 perform on benchmarks?

Sonnet 5 placed fifth in the Intelligence Index with a score of 53; higher-scoring models include GPT-5.5 (xhigh) at 55, Opus 4.7 at 54, Opus 4.8 at 56, and Claude Fable 5 at 60. It improved over Sonnet 4.6 across several evaluations: a six-point overall gain, a 9-point jump on Terminal-Bench v2.1, 10 points on Humanity's Last Exam, and 7 points on SciCode.

On frontier reasoning tests Sonnet 5 still lagged larger or differently configured models. On CritPt, a physics reasoning test from Argonne National Labs and the University of Illinois, Sonnet 5 scored 17 percent, 14 points above Sonnet 4.6 but below GLM-5.2, Claude Opus, Fable, and GPT-5.5 in their higher configurations. Other evaluations stayed roughly flat versus the prior generation.

Why does Sonnet 5 cost more per task despite unchanged token prices?

On paper Sonnet 5 keeps the same token prices as Sonnet 4.6: $3 per million input tokens and $15 per million output tokens. Artificial Analysis found the average task cost for Sonnet 5 to be $2.29, higher than Opus 4.8's roughly $1.97, because Sonnet 5 consumes many more tokens to reach that performance level.

At the maximum performance setting Sonnet 5 uses about 40 percent more output tokens per task than Sonnet 4.6. In agent-based knowledge work benchmarks such as AA-Briefcase and GDPval-AA the model runs about three times as many agent loops as its predecessor. Those behaviors double the practical bill: Sonnet 4.6 cost about $1.20 per task in the same index, while Sonnet 5 rises to $2.29 per task even though the published per-token rates did not change.

Anthropic is offering promotional token rates of $2 or $10 per million tokens through September 1, but Artificial Analysis based its cost comparisons on the regular prices. The effective price increase echoes an earlier change when Opus 4.7 launched with a new tokenizer that split the same text into "approximately 30%" more tokens. Independent measurements found increases of 1.325x to 1.47x, and a community analysis of over 483 submissions recorded a 37.4 percent jump in tokens per request. With Sonnet 5, the tokenizer effect is compounded by the model's more agentic behavior, which consumes additional tokens per task.

Why it matters

Hidden cost creep breaks a basic buying comparison: raw token prices no longer map reliably to real-world bills. Customers comparing models on listed $/million-token rates will miss how generation-to-generation changes in tokenization and agent loops change task economics. That gap makes lower-cost alternatives more competitive: the story notes Deepseek V4 Pro and GLM-5.2 offer competitive performance at a fraction of the cost in the mid-range segment where Sonnet sits.

Vendors who only publish token prices shift the burden of discovery to third-party testing and community analysis. Buyers seeking predictable budgets for knowledge work will favor models or vendors that report cost per standardized task rather than only token rates.

What to watch

Monitor Artificial Analysis and community submissions for updated per-task cost comparisons and token-count measurements as Sonnet 5 sees broader use. Also watch Anthropic's promotional pricing window through September 1 and any further tokenizer or model behavior changes that would alter real cost per job.