Claude Sonnet 5 delivers improved agentic performance at higher cost per task
Anthropic's Claude Sonnet 5 offers enhanced agentic capabilities and improved benchmark performance at a higher cost per task due to increased token usage.
1. Performance and Benchmarking
As of June 30, 2026, Claude Sonnet 5 holds the fifth position on the Artificial Analysis Intelligence Index. The model achieved a score of 53, placing it within 2 to 3 points of GPT-5.5 and Claude Opus 4.8. While Sonnet 5 shows significant improvements over its predecessor, Sonnet 4.6—including gains on benchmarks like Terminal-Bench v2.1, Humanity’s Last Exam, and SciCode—it continues to trail Opus 4.8 in heavy reasoning and knowledge-intensive tasks, such as the CritPt physics benchmark.
2. Agentic Capabilities and Effort
Claude Sonnet 5 demonstrates a shift toward more intensive processing to achieve higher performance. When utilizing the "max effort" setting, the model uses approximately 40% more output tokens per task than Sonnet 4.6 and requires roughly three times the agentic turns for knowledge work evaluations like AA-Briefcase and GDPval-AA. Despite these higher resource requirements, Sonnet 5 matches or slightly outperforms Opus 4.8 in agentic knowledge work tasks. The model features five effort levels, including a new "xhigh" setting, and maintains a 1-million-token context window.
3. Cost and Pricing Structure
The cost per task for Claude Sonnet 5 is approximately $2.29, representing a 15% increase over Claude Opus 4.8 and a twofold increase compared to Sonnet 4.6. This rise in cost is attributed entirely to the model's increased token usage during complex tasks. While the standard pricing remains $3 per million input tokens and $15 per million output tokens, Anthropic is providing a promotional rate of $2 and $10, respectively, until September 1, 2026. Cache pricing remains consistent with previous models, featuring a 25% premium for writes and a 90% discount for hits.