Anthropic releases Claude Sonnet 5 with improved agentic capabilities

After a long wait, Anthropic has finally unveiled Claude Sonnet 5, the newest addition to its mid‑tier lineup. This successor to Sonnet 4.6 skips the 4.x naming convention and jumps straight to 5, aligning its branding more closely with the powerful Opus 4.8 and the popular Fable 5 model.

Concrete real‑world performance data for Sonnet 5 remains scarce, so we must rely on Anthropic’s own announcement as the primary source for now. The company asserts that Sonnet 5 significantly narrows the performance gap with Opus 4.8 while maintaining a markedly lower price point. As expected, the new model outperforms Sonnet 4.6 in reasoning, agentic tasks, tool usage, coding, and general knowledge tasks.

According to Anthropic’s published benchmarks, Sonnet 5 does not surpass Opus 4.8 across every metric. For instance, in agentic coding it achieves 63.2 % versus Sonnet 4.6’s 58.1 %, yet still trails Opus 4.8’s 69.2 %. The only exception is a knowledge‑work benchmark where Sonnet 5 slightly edges out Opus 4.8. For most advanced use cases—including coding—Opus 4.8 remains the superior choice available to the public.

Anthropic also highlights that Sonnet 5 exhibits a reduced incidence of undesirable behaviors compared to Sonnet 4.6, making it safer for agentic applications. This translates to fewer hallucinations and more consistent outcomes when the model operates in agentic contexts.

Today marks the global rollout of Sonnet 5, which is now the default offering for Free and Pro tiers, and is also accessible to Max, Team, and Enterprise subscribers. The new model can be accessed via Claude Code and the Claude API as well.

Sonnet 5 comes with a surprisingly reduced cost compared to its predecessor, Sonnet 4.6—priced at just $2 per million input tokens and $10 per million output tokens. This discounted rate is temporary, lasting only until August 31, after which the price will climb to $3 and $15 respectively.