Feature Request: Reduce Haiku usage cost for Max subscribers in Claude Code

Problem

Max subscribers ($200/month) using Claude Code with agent systems face significant usage pool drain when delegating micro-tasks to subagents. Currently, all model calls (including lightweight validation, hook prompts, and simple formatting tasks) consume the usage pool at the same rate relative to their per-token cost.

For agent-heavy workflows, this means:

• Hook prompts (safety checks, command validation) — already use Haiku, but still consume pool
• Subagent delegation — lightweight research/validation tasks could use Haiku instead of Sonnet, saving 3x pool usage
• Formatting/parsing tasks — simple structured output that doesn't need Sonnet-level reasoning

Proposal

Several options (any would be valuable):

1. Haiku-free tier for Max: Include a generous Haiku allowance (e.g., 50M tokens/month) that doesn't count against the usage pool — similar to how some API plans have free tiers for smaller models
2. Reduced pool consumption: Haiku calls consume pool at a discounted rate (e.g., 0.5x instead of proportional cost)
3. Haiku model selection in Agent tool: Allow model: "haiku" in the Agent tool for Claude Code, letting users explicitly choose Haiku for lightweight subagent tasks (currently blocked for some users)
4. Separate Haiku budget: A dedicated Haiku budget independent of the main Opus/Sonnet pool

Use Cases

• Agent compliance systems: Pre/post tool-use hooks that validate safety, check permissions, verify outputs — these run on every tool call and are perfect for Haiku
• Codebase exploration: Quick file searches, grep operations, simple read-and-summarize tasks that don't need Sonnet reasoning
• Validation subagents: Syntax checking, format validation, simple yes/no decisions
• Multi-agent orchestration: When Opus orchestrates 3-4 Sonnet subagents, some could be Haiku for 3-15x cost reduction

Business Case

• Max subscribers are Anthropic's highest-value individual customers
• Enabling efficient Haiku usage increases what Max users can accomplish per month → higher satisfaction → better retention
• Differentiates Max from Pro more strongly (Pro gets limited Haiku, Max gets generous/free Haiku)
• Encourages sophisticated agent workflows that showcase Claude's capabilities
• Users who hit pool limits early may downgrade or reduce usage — Haiku efficiency keeps them engaged

Current Workaround

Hook prompts already use claude-haiku-4-5 model in settings.json prompt hooks, which works well. But the Agent tool's model parameter doesn't include Haiku as an option in Claude Code, preventing users from choosing the most cost-effective model for simple subagent tasks.

Environment

• Claude Code CLI (Max subscription)
• Heavy agent orchestration workflow (Opus → Sonnet subagents)
• 20-40 hook calls per task, 3-8 subagent calls per task