Claude Sonnet 5 Is Stronger Yet Costlier—Per‑Task Cost Beats Opus 4.8

Anthropic has just launched Claude Sonnet 5, which achieved a score of 53 on the Artificial Analysis intelligence index, 6 points higher than the previous Sonnet 4.6 and tying the high‑reasoning mode of GPT‑5.5.

The higher cost per task ($2.29, 15 % above Opus 4.8) is not caused by a price increase in the API (still $3/$15 per million tokens). Instead, Sonnet 5 consumes about 40 % more output tokens because it “works harder,” and on knowledge‑work benchmarks such as AA‑Briefcase and GDPval‑AA it uses roughly three times more agentic interaction rounds than Sonnet 4.6.

Core data :

Intelligence index ranking: 5th, behind Claude Fable 5, Opus 4.8, GPT‑5.5 (high), etc.

On agentic knowledge‑work benchmarks, Sonnet 5 matches or slightly exceeds Opus 4.8.

On pure reasoning benchmarks (e.g., CritPt physics), it still trails Opus 4.8 and GPT‑5.5.

Context window remains at 1 million tokens; cache pricing unchanged.

Cost comparison :

Anthropic has introduced a promotional price of $2/$10 per million tokens (effective until September 1), roughly a 30 % discount. During the promotion the per‑task cost drops to about $1.60, slightly below Opus 4.8’s standard cost, but the price will revert after the promotion ends.

Token consumption :

Sonnet 5 outputs an average of 69 k tokens per task, second only to GPT‑5.4 mini and nano. For large‑scale inference workloads, this token volume can become costly.

Agentic capability :

On the AA‑Briefcase and GDPval‑AA benchmarks, Sonnet 5 performs on par with or slightly better than Opus 4.8. For agentic applications that require multiple interaction rounds and the generation of professional documents, Sonnet 5 may be the better choice—provided you can afford the token bill.

Detailed evaluation :

Although the Sonnet series has traditionally emphasized “smaller, faster, cheaper,” Sonnet 5 abandons the “cheaper” label to chase higher intelligence. It excels in agentic tasks, but for simple question‑answering or code‑completion scenarios, Opus 4.8 or even Sonnet 4.6 may be more economical.

Anthropic appears to be betting that users will pay extra for stronger agentic abilities, especially as OpenAI’s GPT‑5.6 approaches market launch, intensifying the price competition.

If you are considering migrating to Sonnet 5, run your typical workloads first and calculate token consumption. The promotional period is a good time to test, but don’t be swayed solely by the “stronger” claim—your wallet will speak.

More comparative data: https://artificialanalysis.ai/models/claude-sonnet-5