Cost Optimization

Budget Allocation Strategy

The triad system is designed around a tiered cost model:

Tier	Budget %	Use Case
Composer 1.5	80%	Default for all standard work
Codex Standard	15%	General implementation, tests
Codex High	4%	Security, algorithms, complex logic
Claude (Opus/Sonnet)	1%	Strategy only (via Claude Code)

The 80/15/4/1 Rule

80% Composer: Quick edits, iteration, formatting, simple fixes
15% Codex Standard: Feature implementation, test generation, refactoring
4% Codex High: Security audits, performance optimization, complex algorithms
1% Claude: Strategic design and architectural decisions

This allocation maximizes quality while minimizing cost. Most coding tasks don’t need expensive models.

composer_only Mode

For strict budget control, enable composer_only in etc/cursor_config.json:

{
  "composer_only": true,
  "model_override": "composer-1.5"
}

This forces ALL Cursor invocations to use Composer 1.5, regardless of what model is requested. Use this when:

Approaching monthly credit limits
Doing routine maintenance work
Running many parallel tasks
Budget is the primary concern

Credit Monitoring

Check Current Model

cat etc/cursor_config.json | python3 -c "import json,sys; d=json.load(sys.stdin); print(f'Model: {d[\"default_model\"]}, Composer Only: {d.get(\"composer_only\", False)}')"

Budget Configuration

{
  "credit_target_monthly": 500,
  "budget": {
    "allocation": {
      "composer": 0.80,
      "codex_standard": 0.15,
      "codex_high": 0.04,
      "codex_xhigh": 0.01
    }
  }
}

Cost Savings Techniques

1. Use the Lowest Sufficient Tier

Task	Tier	Why
Fix typo	Composer 1.5	Trivial change
Add function	Codex Standard	Needs understanding
Security audit	Codex High	Needs deep analysis
Variable rename	Composer 1.5	Simple edit
Algorithm design	Codex High	Complex reasoning

2. Bridge Agents Use Haiku

Bridge agents (the “dispatch” leg of the triad) use Claude Haiku, which is very cheap. They don’t do any real work — they just route CLI commands.

3. Batch Similar Tasks

Instead of:

cursor-agent "Fix typo on line 10"
cursor-agent "Fix typo on line 20"
cursor-agent "Fix typo on line 30"

Do:

cursor-agent "Fix typos on lines 10, 20, and 30"

4. Use Profiles for Common Tasks

# Quick edit profile uses composer-1.5 automatically
cursor-agent -p quick-edit "Rename variable x to count"

5. Dry-Run Before Expensive Operations

# Check routing without executing
./bin/subagent-router --dry-run "Audit security of auth module"
# Output: Would route to: cursor (tier: codex-high)

When to Upgrade Tiers

Signal	Action
Composer output has bugs	Upgrade to Codex Standard
Logic errors in complex code	Upgrade to Codex High
Security vulnerability found	Mandate Codex High for all security
Simple fix taking too long	Stay on Composer, simplify the prompt
Test failures after generation	Add Composer review step

Disabling Cursor Agents

To completely disable all Cursor agent dispatch:

{
  "enabled": false
}

This causes all cursor-agent invocations to exit with an error, preventing any credit usage.

Cost Comparison Table

Operation	Composer 1.5	Codex Standard	Codex High
Relative cost	1x	~3x	~7x
Speed	Fast	Medium	Slower
Quality	Good	Better	Best
Best for	Edits, iteration	Features, tests	Audits, algorithms