Anthropic ships Claude 3.5 Sonnet at the Opus price point

Anthropic released Claude 3.5 Sonnet, a mid-tier model priced at $3 per million input tokens and $15 per million output. It outperforms the previous flagship, Claude 3 Opus, on graduate-level reasoning (GPQA), undergraduate-level knowledge (MMLU), and HumanEval code generation. Anthropic also shipped 'Artifacts' — a side-panel canvas for rendered output.

The pricing is the most consequential part. Sonnet is 1/5 the cost of Opus for superior benchmark scores. Anthropic is signalling that the cost curve of frontier-quality retrieval is falling fast.

When the best model gets cheaper, AI-assistant surfaces become more ambitious. Claude is now cheap enough that Anthropic's own shopping-assistant experiments, and downstream products built on the Claude API, can afford deeper retrieval per query. Deeper retrieval means more citations per answer — and Shopify merchants cited in those answers sit in the funnel.

Better reasoning also raises the bar for what content makes the citation cut. Claude 3.5 Sonnet is better at discarding weak sources. Catalogs with thin descriptions, missing schema, or hallucination-prone content get filtered out more aggressively. The asymmetric payoff is to the stores whose pages are parseable, factual, and clean.

For merchants, the tactical move is to make sure every product page answers the specific questions buyers ask before buying — materials, fit, comparisons to named alternatives, return policy in one sentence. Claude 3.5 Sonnet rewards that density; thin product pages get skipped.