Anthropic releases Claude Sonnet 5
Claude Sonnet 5 introduces significant improvements in agentic performance, including better reasoning, tool use, and coding capabilities, available at a lower price point than previous models.
Why it matters
It represents a shift in cost-efficiency for agentic tasks, providing performance that nears Opus-class models while remaining accessible for enterprise and developer use cases.
TL;DR
- 01Stronger agentic capabilities for coding and tool use.
- 02Cost-effective performance compared to Opus-class models.
- 03Safer, with lower hallucination and sycophancy rates than Sonnet 4.6.
Performance and Cost
Sonnet 5 performance nears Opus 4.8. Pricing through August 31, 2026, is $2/MTok input and $10/MTok output. After this, it shifts to $3/MTok input and $15/MTok output.
Agentic Benchmarking
The model shows strict improvements over Sonnet 4.6 in BrowseComp and OSWorld-Verified benchmarks. Partners note its ability to finish complex tasks that previously caused agents to stall.
Safety
Evaluations show Sonnet 5 is safer and more resistant to hijacking, but possesses lower capabilities for dangerous cybersecurity tasks than Opus 4.8 or Mythos models.