Anthropic launches Claude Sonnet 5 as cheaper agentic AI near Opus quality

Anthropic has launched Claude Sonnet 5, a midsize model designed to bring high-end agentic capabilities at a lower price point. Priced at $2 per million input tokens and $10 per million output tokens through August 31 (rising to $3/$15 thereafter), Sonnet 5 rivals Opus 4.8 in knowledge work and achieves a 63.2% score on agentic coding benchmarks, compared to Opus 4.8's 69.2% and its predecessor Sonnet 4.6's 58.1%. The model excels at planning, tool use (browsers, terminals), and autonomous task execution, matching the agentic push of OpenAI's GPT-5.6 Sol and Google's Gemini 3.5 Flash. Anthropic positions it as a cost-effective alternative that lets developers balance performance and expense.

Sonnet 5 also demonstrates significant safety improvements: lower rates of hallucination, sycophancy, and undesirable behaviors like cooperation with misuse. It refuses malicious requests more reliably and resists prompt-injection attacks. Testers from Zapier and Lovable note its ability to complete complex multi-step tasks without stalling, and its clean refusal of unsafe requests. While not as robust as Opus 4.8 on the hardest problems, Sonnet 5 sets a new baseline for affordable agentic AI—making autonomous, reliable agents accessible to a wider range of developers and automating workflows that previously stalled.