Anthropic released Claude Sonnet 5 on June 30, 2026, calling it the most agentic model yet in its mid-tier Sonnet line. The company says the update narrows the performance gap with its flagship Opus 4.8 model while keeping costs closer to a mainstream tier, aiming the release at businesses running AI agents that plan tasks and operate tools such as browsers and terminals with minimal supervision.
What's new
According to Anthropic, Sonnet 5 improves on its predecessor, Sonnet 4.6, across reasoning, tool use, coding, and general knowledge work, and can complete multi-step tasks autonomously at a level that previously required larger models. The company also reports a lower rate of what it calls "undesirable behaviors" in agentic settings compared with Sonnet 4.6, alongside reduced cybersecurity-relevant capabilities and safeguards enabled by default.
Pricing undercuts Anthropic's own top-tier model: Sonnet 5 launches at $2 per million input tokens and $10 per million output tokens through August 31, rising to $3 and $15 afterward. TechCrunch reported the pricing positions Sonnet 5 as cheaper than Opus 4.8, OpenAI's GPT-5.5, and Google's Gemini 3.1 Pro.
Sonnet 5 became the default model for Anthropic's Free and Pro plans on release, with access also extended to Max, Team, and Enterprise tiers, the Claude Platform API, and Claude Code.
Industry context
The launch arrived amid a broader shift among frontier labs toward multiple model tiers built for different budgets and speeds rather than a single flagship release, following recent updates from OpenAI and Google. Early users cited by TechCrunch, including engineers at Zapier and Lovable, described gains in completing longer automation workflows without stalling partway through, though both also noted that reliability, not just raw capability, is now the main differentiator among competing agentic models.
Anthropic has not disclosed independent third-party benchmark comparisons beyond its own system card; claims of improved safety and reduced misuse risk are based on the company's internal evaluations.
Source: TechCrunch