Agentic AI
Anthropic Rolls Out Claude Sonnet 5 With Agent Smarts at Lower Costs
The midsize model brings agentic capabilities—autonomous planning and tool use—to a cheaper price tier, intensifying competition in the AI race to democratize agent technology.
TechCrunch — AI
Anthropic has unveiled Claude Sonnet 5, a refreshed version of its midsize model designed to handle autonomous agent tasks at a fraction of the cost previously required. The release underscores a shift in how foundation model makers are competing: as agent capabilities become standard across the industry, the battleground is moving from capability to price and reliability.
The Agentic Shift#
The new model can plan, execute tools like browsers and terminals, and operate independently—abilities that until recently demanded Anthropic's larger, more expensive options. This mirrors recent moves by rivals: OpenAI launched GPT-5.6 Sol in preview with subagent support, while Google's Gemini 3.5 Flash pivoted toward autonomous work in May. Anthropic's framing suggests the window where companies compete on who can do agent work is closing; the real differentiation now lies in affordability and autonomy without constant human intervention.
Pricing and Availability#
Sonnet 5 launches as the default model for free and Pro plans, with tiered pricing designed to undercut competing offerings. Through August 31, it costs $2 per million input tokens and $10 per million output tokens—rates that will rise to $3 and $10 respectively after the promotion expires. This places it cheaper than both Opus 4.8 and OpenAI's GPT-5.5, though pricier than Gemini 3.5 Flash. The model becomes available starting Tuesday across all subscription tiers.
Performance Gains#
Sonnet 5 delivers performance approaching Opus 4.8 (Anthropic's heavyweight) while substantially improving over Sonnet 4.6 (February's release). On a coding benchmark, it scores 63.2%—below Opus 4.8's 69.2% but well ahead of Sonnet 4.6's 58.1%. More notably, on knowledge work tasks, Sonnet 5 slightly outperforms Opus 4.8, hinting that mid-tier pricing doesn't necessarily mean mid-tier reasoning on complex judgment calls.
Testers reported the model excels at completing multi-step workflows without stalling midway and can validate its own output unprompted. Daniel Shepard, senior engineer at Zapier, noted in a statement that Sonnet 5 "finished end to end" on a two-part task—updating Salesforce tiers and sending announcements—that would have previously "stall[ed] halfway."
