Feature: Per-tool model routing to optimize cost and latency

Problem

When using Claude Code with expensive models (Opus 4.6), every tool call — including simple, mechanical ones — runs on the same high-cost model. Many tools don't need the reasoning power of a frontier model:

• Playwright/Browser tools (screenshot, snapshot, click, navigate) — these are mechanical interactions that just need to follow instructions, not reason deeply
• File reads — reading a file doesn't require advanced reasoning
• Simple bash commands — running git status, ls, npm install, etc.
• MCP tool calls — many MCP tools just fetch data or perform simple actions

Currently, the only options are:

1. Use an expensive model for everything (wasteful for simple tool calls)
2. Use a cheap model for everything (insufficient for complex reasoning)
3. Manually switch models mid-session with /model (tedious and breaks flow)

Proposed Solution

Allow users to configure model routing per tool or tool category in settings.json or .claude/settings.json:

{
  "model": "opus",  // default model for reasoning/planning
  "toolModels": {
    // Route specific tools to cheaper/faster models
    "mcp__playwright__*": "haiku",     // all Playwright tools
    "Bash": "sonnet",                   // bash commands
    "Read": "haiku",                    // file reads
    "Glob": "haiku",                    // file searches
    "Grep": "haiku",                    // content searches
    "mcp__figma-remote-mcp__*": "sonnet" // Figma tools
  }
}

How it would work

1. The main conversation uses the primary model (e.g., Opus) for reasoning, planning, and code generation
2. When a tool call is made, if the tool matches a toolModels pattern, the response processing for that tool's output uses the specified cheaper model
3. The context/conversation history remains unified — only the model used for processing tool results changes

Alternative: Tool category presets

{
  "toolModelPresets": {
    "browser": "haiku",      // all browser/Playwright tools
    "filesystem": "haiku",   // Read, Glob, Grep, Write, Edit
    "shell": "sonnet",       // Bash
    "mcp": "sonnet"          // all MCP tools
  }
}

Why This Matters

• Cost: Opus is ~5x more expensive than Sonnet and ~60x more than Haiku. A session with heavy browser testing or file exploration can burn through credits on tool calls that don't benefit from Opus-level reasoning
• Latency: Cheaper models respond faster. Browser interaction loops (click → snapshot → click) would be significantly snappier with Haiku
• No quality loss: A screenshot inspection, file read, or git status execution produces the same result regardless of which model processes it

Current Workarounds (all suboptimal)

| Workaround | Downside |
|---|---|
| /model sonnet before tool-heavy work | Manual, breaks flow, easy to forget to switch back |
| CLAUDE_CODE_SUBAGENT_MODEL | Only applies to Task subagents, not direct tool calls |
| Just use Sonnet for everything | Loses Opus reasoning quality for complex tasks |

Environment

• Claude Code CLI
• Primary model: Opus 4.6
• Tools that would benefit from routing: Playwright MCP, Figma MCP, filesystem tools, Bash