Why We Stopped Letting Premium Models Plan — And Cut AI Coding Bills 70%

TL;DR: A Cursor agent burned $1,400 in one hour with zero output — 40% of tokens were planning at $30/M when $0.25/M models do it better. Full comparison →

The Context

We run a small dev team (3 engineers) using Cursor and Claude Code daily. Last month a background agent looped for an hour, consumed 1.3B tokens, and produced nothing — the CEO refunded it. That bill was a wake-up call: we had zero task-level brakes, no per-seat caps, and no routing logic. Most teams are in the same spot — paying Opus/GPT-4o rates for tokens that just “think about how to think.”

What We Tested

Tool	Use Case	Verdict	Why
Cursor (default)	Agent runs with Opus/Sonnet for everything	❌	No routing, no budgets, no task-level limits
LangSmith	Org/workspace/user/key budget controls + dynamic pricing	✅	Only platform with enforceable per-seat daily caps and routing rules
Custom token router (Haiku → Sonnet)	Planning → cheap, execution → premium	✅	60-70% cost reduction, same output quality
Cursor Origin (announced)	AI-native git + agent orchestration	?	Fall 2026 — promises built-in routing, unproven

The Pivot Point

The $1,400/hr runaway loop. We realized 40% of tokens were planning — deciding file paths, refactor strategies, test approaches — all billed at $15-30/M. A Haiku call at $0.25/M produces equally useful plans. The waste wasn’t marginal; it was structural.

What We Use Now

A lightweight YAML router sitting in front of our agent runs:

routing:
  planning: claude-3-haiku-20240307
  execution: claude-3-5-sonnet-20241022
  budgets:
    daily_per_seat: 500000  # tokens
    task_max: 100000
    alert_threshold: 0.8

LangSmith enforces the budgets; the router handles model selection. Our monthly AI coding spend dropped from ~$2,800 to ~$850.

When You’d Choose Differently

If you’re a solo dev doing <5 agent runs/day, the router overhead isn’t worth it — just set a manual daily cap in Cursor and check bills weekly. Also, if your planning requires massive context (entire codebase + docs), Haiku’s 200K window may truncate; Sonnet 3.5’s 200K+ with better reasoning wins there.

Tool Crucible Rating

Dimension	Rating
Overall	★★★★☆
Ease of Setup	★★★☆☆
Value (Cost Savings)	★★★★★
Support/Documentation	★★★☆☆

---

This is part of our AI Coding Cost Control evaluation series. See full comparison: AI Coding Tool Cost Breakdown 2026