Anthropic's Claude Sonnet 5 Closes the Gap on Opus at a Fraction of the Price

Anthropic has released Claude Sonnet 5, positioning it as its most agentic mid-tier model to date. The pitch is that work which recently demanded larger, pricier Opus-class models — multi-step planning, browser and terminal use, sustained autonomous execution — now runs on a cheaper Sonnet. Anthropic claims performance approaching Opus 4.8 while undercutting it on cost, with API pricing starting at an introductory $2 per million input tokens and $10 per million output through August 31, 2026, before settling at $3 and $15. The model is the new default on Free and Pro plans and is available across Claude Code, the API, and enterprise tiers.

Early-access partners emphasize follow-through on real engineering tasks: tracing bugs to root cause rather than patching symptoms, writing reproducing tests, verifying its own fixes, and carrying messy pull requests through to a tested result without hand-holding. Anthropic highlights an effort dial that lets developers trade cost against accuracy, with Opus 4.8 still recommended when maximum quality matters. Reported gains span coding, tool use, reasoning, and knowledge work over the prior Sonnet 4.6.

On safety, Anthropic says Sonnet 5 refuses malicious requests more reliably, resists prompt-injection hijacking better, and shows lower hallucination and sycophancy than its predecessor — though it remains less aligned than Opus 4.8 on internal behavioral audits. Notably, the model was not trained for offensive cybersecurity and never produced a working exploit in evaluations, but its general-intelligence gains nudged up partial-exploit rates enough that Anthropic shipped it with real-time cyber-abuse safeguards enabled by default, the same controls used in recent Opus releases.