Introducing Claude Sonnet 5
02:00 · June 30, 2026 · Anthropic News

Summary
Anthropic has released Claude Sonnet 5, positioning it as the most capable agentic model in the Sonnet line to date. The model plans multi-step actions, invokes tools such as browsers and terminals, and sustains autonomous operation on tasks that previously demanded larger Opus-class systems. It narrows the performance difference with Opus 4.8 while remaining substantially cheaper, delivering measurable gains over Sonnet 4.6 in reasoning, tool use, coding, and knowledge-work workflows.
Benchmarks on agentic search (BrowseComp) and computer-use (OSWorld-Verified) evaluations show Sonnet 5 outperforming its predecessor across effort levels and offering a broader cost-performance range than Opus 4.8. Users can tune effort settings to balance expense against results, with medium-effort operation providing particularly strong efficiency and higher-effort runs approaching Opus 4.8 on selected tasks.
Safety evaluations indicate lower overall rates of undesirable behaviors compared with Sonnet 4.6, including improved resistance to prompt-injection hijacking, reduced hallucination and sycophancy, and safer refusal of malicious requests. On cybersecurity benchmarks the model performs routine, non-harmful tasks but shows markedly weaker results than Opus 4.8 or Mythos 5 at developing working exploits; default real-time safeguards block dangerous cyber usage.
Early-access partners report that the model completes complex, multi-step assignments where earlier Sonnet versions halted prematurely, performs self-verification without explicit prompting, and maintains these capabilities at lower cost. Sonnet 5 is now the default model for Free and Pro plans, available to all other tiers, and accessible via the Claude API at introductory rates of $2 per million input tokens and $10 per million output tokens through August 2026, after which standard pricing applies.
Why it matters
Direct model release with actionable performance data, pricing, safety details, and workflow examples for builders implementing agentic AI in production. Specific versions, benchmarks, and safeguards enable immediate evaluation and integration decisions.






