Anthropic released Claude Sonnet 5 on 30 June, describing it as its most agentic Sonnet model. It could plan, use browsers and terminals, write code, and handle professional work at a lower listed price than the larger Opus tier.
Why the Sonnet tier mattered
For an agent, token price is only one part of cost. A long task may call several tools, retry steps, create files, and require human review. A cheaper model can save money, but it can also encourage teams to automate more tasks before they have the controls to supervise them.
Cost of input and output tokens.
Tools, retries, elapsed time, and review.
The cost of an incorrect or unauthorised action.
What you can try
Compare Sonnet 5 and the current alternative on one complete workflow. Measure successful completion, unsafe attempts, tool calls, time, token use, and human correction. Run the test more than once because agent results can vary.
Keep the result easy to verify.
Open only required tools and destinations.
Record both success and unsafe attempts.
Include review and recovery work.
Key takeaways
- A spending limit stops runaway work.
- Sensitive actions require approval.
- Activity logs are retained.
- The team can disable the agent quickly.
Sonnet 5 made capable agents more accessible. That strengthened the case for permission limits, test accounts, checkpoints, and spending controls before wider deployment.
