Anthropic’s latest Sonnet release gives teams a more practical question than ‘which model is best?’ The question is where higher effort changes the result enough to justify the extra time, tokens and review overhead.
In its Claude Sonnet 5 announcement, Anthropic says the model is available across Claude plans, Claude Code and the Claude Platform, with introductory API pricing through 31 August 2026. The company also positions the model around adjustable effort levels, cost-performance tradeoffs and stronger agentic search and computer-use behaviour.
For small teams, that should change testing practice. A higher-effort model setting should earn its place in a workflow. It should not become the default because the release page sounds impressive.
Start with tasks where follow-through matters
Anthropic describes Sonnet 5 as stronger at multi-step agent work, including search, tool use, coding and computer-use tasks. That points to a specific test set: jobs where a model has to keep context, use tools in order, check its own result and stop safely when the instructions run out.
Good candidates include triaging support inboxes, turning a product brief into tracked implementation tasks, preparing a first pass on a code change, or checking a campaign report against source data. Weak candidates include one-line rewriting, simple classification and short extraction jobs. Those usually do not need more expensive reasoning.
Kahunam has covered similar testing discipline in Gemini 3.5 and action-first AI: what small teams should test before automating more work. The same principle applies here: test the workflow, not the model name.
Measure more than completion
A task can complete and still be a poor automation candidate. Track whether the output is correct, whether the model used the right source, whether it respected permissions, whether a human can review the result quickly, and whether the same task succeeds again with slightly different inputs.
For each test, run the same brief at two or three effort levels. Record elapsed time, token cost, tool calls, errors, required human edits and the final decision. If higher effort mainly produces longer explanations, keep the lower setting. If it catches missing context, creates a better verification step or avoids a failed handoff, the extra cost may be justified.
Separate agent work from approval work
Anthropic says Sonnet 5 has cyber safeguards enabled by default and refers readers to the model system card for safety and capability evaluations. That does not remove the need for local rules. The model can draft, inspect and recommend. It should not quietly approve its own work where money, publishing, customer access or production systems are involved.
Set a clear boundary for every workflow before testing: what the agent may read, what it may change, what it must ask about, and what evidence it must leave behind. A stronger model is useful when it follows those boundaries more consistently. It is risky when the process treats extra capability as permission to skip review.
A simple pilot plan
Pick five repeatable tasks that already consume team time. Write one reusable brief for each. Run each task at low, medium and high effort, then score the output against accuracy, source use, review time, cost and repeatability. Keep the winning configuration only where the business case is visible.
The practical takeaway is straightforward: Claude Sonnet 5 gives teams a reason to refresh agent tests while keeping automation scoped. Start where follow-through matters, measure the tradeoff, and keep human approval around decisions that can affect customers, spend or live systems.
Author
