Anthropic releases Claude Sonnet 5 as default agentic model
The new mid-tier model prioritizes autonomous terminal and browser use, though tokenizer changes have led to higher effective costs in initial benchmarks.
Julian Reeve
Jul 3, 2026 · 1 min read
One million tokens of context now serve as the new baseline for Anthropic’s mid-tier flagship. Claude Sonnet 5 arrived this week, positioned as the company’s most “agentic” model to date, designed specifically for autonomous execution in browsers and terminals. The release includes a promotional pricing tier through late summer, setting rates at $2 per million input tokens and $10 per million output tokens, undercutting previous frontier release prices.
While the model shows a meaningful performance increase on coding benchmarks—scoring 57% on CursorBench compared to its predecessor's 49%—the rollout has been met with scrutiny regarding its efficiency. Third-party analysis suggests that tokenizer changes and a higher frequency of “turn-taking” in agentic tasks can make Sonnet 5 more expensive to run than some existing high-end models for specific workloads. Despite these cost dynamics, the model has seen immediate integration across developers like Cursor, Perplexity, and Devin.
The launch also serves as a platform update. Anthropic has introduced a Linux beta for its desktop client and updated its managed agents with streaming session deltas and credential injection scoping. Notably, the anticipated release of the more powerful “Fable 5” remains pending government approval, leaving Sonnet 5 as the current primary default for the Anthropic ecosystem. Security remains a focus, with cyber safeguards enabled by default, though the company notes that its Opus tier remains the preferred choice for high-stakes vulnerability research.