Claude Sonnet 4.6: Effort and Tool-Triggering Configuration

One-sentence takeaway

Sonnet 4.6's effort affects text, thinking, and tool-call tokens at the same time; Anthropic recommends that most production applications start at medium and set it explicitly to avoid the unpredictable latency caused by the default high setting.

Use cases

• Suitable tasks: Tuning the cost, latency, and number of tool calls for Claude APIs/agents, as well as multi-turn coding and enterprise document tasks.

• Unsuitable tasks: Treating low as “the same quality” without evals; it is also unsuitable to use max for every short Q&A.

• Applicable model version: Claude Sonnet 4.6, with the model string claude-sonnet-4-6.

• Applicable clients, agents, or APIs: Supported platforms such as the Claude API, AWS Bedrock, Google Cloud, and Microsoft Foundry; use the request fields required by the target platform.

• Recommended reasoning tier and parameter: medium is the recommended default; use low for low-latency or non-coding chat, high for complex reasoning or code, and max only when evals show a benefit. Sonnet 4.6 supports max, but xhigh is not listed as supported in the official table.

Ready-to-use content

API configuration skeleton (field names follow Anthropic's documentation; verify the SDK version before deployment):

{
  "model": "claude-sonnet-4-6",
  "max_tokens": 8192,
  "output_config": {
    "effort": "medium"
  },
  "system": "Use necessary tools to complete the task; do not provide suggestions only. Verify the result after every write; if information is insufficient, say so explicitly.",
  "tools": [
    {
      "name": "search_docs",
      "description": "Search the approved corpus and return source IDs. Use when facts are missing or may have changed.",
      "input_schema": {
        "type": "object",
        "properties": {"query": {"type": "string"}},
        "required": ["query"]
      }
    },
    {
      "name": "write_file",
      "description": "Write only the requested file path, then return its path and a verification summary.",
      "input_schema": {
        "type": "object",
        "properties": {"path": {"type": "string"}, "content": {"type": "string"}},
        "required": ["path", "content"]
      }
    }
  ]
}

The accompanying system prompt should use normal, specific wording about “when to use tools”: Use search_docs when external or user-specific facts are needed; use write_file only after the requested content is complete.

Test/workflow steps

1. Fix the input, tool descriptions, max_tokens, and model snapshot, and scan only effort=low/medium/high/max.

2. Record complete response tokens, thinking tokens (when enabled), the number of tool calls, time to first token, total latency, and task success rate.

3. Test “suggest a modification” and “execute a modification” instructions separately to verify whether explicit action wording changes the tool-trigger rate.

4. If prompt caching is used, keep effort unchanged within the same session; adjust effort only when comparing across workloads.

5. Wrap write tools with permission checks, path allowlists, content review, and read-after-write verification.

Raw evidence and data

• The official effort documentation states that effort affects all tokens in a response, including text, tool arguments, and thinking; lower effort typically means fewer tool calls, more direct action, and shorter confirmations.

• Sonnet 4.6 defaults to high; Anthropic officially recommends medium for most Sonnet 4.6 applications as the balance of speed, cost, and performance, low for high-throughput or latency-sensitive tasks, high for complex reasoning, and max when absolute capability is the priority.

• The official documentation states that xhigh is not in the available list for Sonnet 4.6, while max is available.

• Anthropic advises phrasing tool-trigger instructions as explicit actions rather than “could you suggest a modification”; it also warns that overly forceful MUST/CRITICAL language may cause newer models to trigger tools excessively.

• The official documentation states that changing effort in long conversations that depend on caching invalidates the previously cached prefix; effort should remain stable within a cached session.

Applicability boundaries

• Different cloud platforms may add support for output_config.effort, thinking, or tool fields at different times; use the response from the actual endpoint as the authority.

• Effort controls tokens/workload, not a hard limit on visible answer length; to control length, still be explicit in the prompt and max_tokens.

• Fewer tool calls are not necessarily better; write tasks should prioritize correctness and verification rather than only optimizing the call count.

• There is no evidence that max is “always more correct”; use real business evals to calculate the incremental quality and cost.

Source excerpts or observations (compliance short quote)

Anthropic's recommendation for Sonnet 4.6 is “Medium effort (recommended default)” (a compliant short quote).