# AI Cost Tracker
<!-- Track LLM usage cost per agent/job to avoid billing surprises -->
<!-- Update when: changing model, changing cron frequency, adding new agent -->
<!-- last updated: 2026-03-25 -->

---

## Model Pricing Reference

| Model | Input (per 1M) | Output (per 1M) | Best for |
|-------|---------------|-----------------|---------|
| DeepSeek Reasoner | $0.14 | $0.28 | Most agent tasks, reasoning, strategy |
| Claude Haiku 4.5 | $0.25 | $1.25 | Classification, extraction, simple tasks |
| Perplexity Sonar | $3.00 | $15.00 | Web search, live data |
| Claude Sonnet 4.6 | $3.00 | $15.00 | Vision, complex multi-step, code review |
| MiniMax (image) | per image | — | Image generation (TEDx, RecruitAI) |

---

## Scheduled Jobs — Monthly Cost Estimate

| Job | Schedule | Model | Est. tokens/run | Est. $/month |
|-----|----------|-------|-----------------|-------------|
| Marketing strategy | 1x/day | DeepSeek Reasoner | ~8K in / ~3K out | ~$3.60 |
| Market sentinel | 1x/day | Perplexity Sonar | ~5K in / ~2K out | ~$5.40 |
| Tender digest | 1x/day | DeepSeek Reasoner | ~10K in / ~4K out | ~$4.50 |
| Topic intelligence | 1x/day | Perplexity Sonar | ~6K in / ~2K out | ~$6.00 |
| **Total scheduled** | | | | **~$19.50/mo** |

> ⚠ Known mistake: marketing strategy was accidentally set to 5x/day = $18/mo alone. Always verify schedule before deploying.

---

## On-Demand Agents — Per-Call Cost Estimate

| Agent | Model | Est. tokens/call | Est. $/call |
|-------|-------|-----------------|------------|
| brandStrategyAgent | DeepSeek Reasoner | ~15K in / ~5K out | ~$0.004 |
| researchAgent | Perplexity Sonar | ~8K in / ~3K out | ~$0.069 |
| competitorAgent | Perplexity Sonar | ~8K in / ~3K out | ~$0.069 |
| seoAgent | DeepSeek Reasoner | ~10K in / ~4K out | ~$0.003 |
| socialAgent | DeepSeek Reasoner | ~10K in / ~4K out | ~$0.003 |
| customerInsightAgent | DeepSeek Reasoner | ~12K in / ~5K out | ~$0.003 |
| leadScoringAgent | Claude Haiku | ~3K in / ~1K out | ~$0.002 |
| ziweiInterpretationAgent | DeepSeek Reasoner | ~20K in / ~8K out | ~$0.005 |
| csoOrchestrator | DeepSeek Reasoner | ~30K in / ~10K out | ~$0.007 |
| smeGrowthOrchestrator | DeepSeek Reasoner | ~25K in / ~8K out | ~$0.005 |

---

## Model Selection Rules

```
Task type                     → Model
─────────────────────────────────────────────────
Needs live web data           → Perplexity Sonar
Simple extraction/classify    → Claude Haiku
Everything else               → DeepSeek Reasoner
Has image input               → Claude Sonnet
Complex multi-model workflow  → Claude Sonnet
```

**Cost check before changing model:**
- Switching from DeepSeek → Perplexity = ~20x cost increase
- Switching from DeepSeek → Claude Sonnet = ~20x cost increase
- Always estimate monthly impact before changing a scheduled job's model

---

## Budget Alerts

Update this section when you set/change any budget guardrails:

| Alert | Threshold | Action |
|-------|-----------|--------|
| Scheduled jobs total | > $30/mo | Review frequencies |
| Single agent call | > $0.50 | Check token bloat, truncate context |
| MiniMax images | > $20/mo | Review generation triggers |

---

## Cost Change Log

| Date | Change | Before | After | Reason |
|------|--------|--------|-------|--------|
| 2026-02-01 | Marketing strategy 5→1x/day | ~$18/mo | ~$3.60/mo | Over-scheduled by mistake |