Anthropic launches Claude Sonnet 5: agentic at near-Opus quality, $2/$10 intro pricing

Anthropic released Claude Sonnet 5 on June 30, 2026, positioning it as the most agentic model in the Sonnet line. The model is now the default for Free and Pro users and is available across all Claude platforms, including the API and Claude Code. Introductory pricing is $2 per million input tokens and $10 per million output tokens through August 31, 2026; standard pricing from September 1 is $3 input and $15 output—identical to Sonnet 4.6's rate card despite significantly stronger agentic, coding, and tool-use performance.

Sonnet 5 narrows the performance gap to the flagship Opus 4.8 model on many benchmark tasks: on knowledge work (GDPval-AA v2) Sonnet 5 edges ahead at 1,618 vs. Opus 4.8's 1,615; on agentic search (BrowseComp) and computer-use tasks (OSWorld-Verified), Sonnet 5 demonstrates strong multi-step execution, planning, and tool coordination. Testers report that Sonnet 5 completes autonomous multi-step workflows (e.g., CRM updates + email dispatch) that previously required Opus-tier models or human intervention.

Important caveat: Anthropic introduced an updated tokenizer for Sonnet 5 that produces roughly 1.0– 1.35× more tokens from the same text depending on content type. The intro pricing ($2/$10) is set to keep the cost-per-task roughly neutral versus Sonnet 4.6, but from September 1 onward, the standard rate ($3/$15) plus the tokenizer effect can raise effective costs 20–35% above baseline for some workloads despite the nominally unchanged sticker price.

For infrastructure and ML teams, Sonnet 5 is strategically significant as a cost-performance inflection: agentic capability—long reserved for premium tiers—now reaches mid-market pricing with production-grade tool use and autonomous task completion. The introduction signals that multi-step, autonomous workflows are becoming table-stakes across model tiers, matching trends in OpenAI's GPT-5.6 Sol and Google's Gemini 3.5. Teams should measure token consumption on real workloads before September 1 to budget accurately past the intro window.