# Claude Models - Sonnet + Haiku Only (2026-02-17)

## Configuration

**Primary Model:**

- Claude Sonnet 4.5 (`anthropic/claude-sonnet-4-5`)
- Via Claude Max proxy (localhost:3456)
- Cost: €0 (included in subscription)

**Fallback Chain:**

1. Claude Haiku 4.5 (via Claude Max proxy) - Fast + cheap
2. GPT-5 Mini (GitHub Copilot) - Included in $20/month
3. DeepSeek V3.2 (Nvidia) - Free tier
4. Gemini 3 Flash (Google) - Free tier

**Removed:**

- ❌ Claude Opus 4.5 (too expensive, even via proxy)

## Cost Comparison (via Claude Max Proxy)

| Model          | Speed  | Quality | Usage Strategy          |
| -------------- | ------ | ------- | ----------------------- |
| **Sonnet 4.5** | Medium | High    | Primary (general tasks) |
| **Haiku 4.5**  | Fast   | Good    | Fallback (quick tasks)  |
| ~~Opus 4.5~~   | Slow   | Best    | ❌ Removed (overkill)   |

All Claude usage now counts against €90/month Claude Max subscription (no additional API charges).

## Why This Setup?

1. **Sonnet** handles 95% of tasks perfectly
2. **Haiku** is 10x faster for simple tasks
3. **Opus** is overkill and uses more subscription quota
4. **Fallbacks** provide redundancy if Claude quota runs out

## Total Monthly Cost

- Claude Max: €90 (Sonnet + Haiku, unlimited)
- API charges: **€0** ✅
- GitHub Copilot: $20 (GPT-5 Mini fallback)
- Total: ~€110/month

**Previous cost:** ~€1,700/month (with API charges)  
**Savings:** ~€1,590/month! 🎉

## Usage Distribution (Expected)

- 90% Sonnet (primary)
- 5% Haiku (quick tasks, if quota limits)
- 5% Other fallbacks (rare, if Claude unavailable)

## Management

**Check current model:**

```bash
openclaw config get agents.defaults.model
```

**Switch to Haiku temporarily:**

```bash
openclaw config set agents.defaults.model.primary "anthropic/claude-haiku-4-5"
openclaw gateway restart
```

**Check Claude Max usage:**

- Visit https://claude.ai/settings/limits
- Monitor message counts (not token counts)

## Notes

- Both Sonnet and Haiku go through proxy (localhost:3456)
- Proxy uses OAuth token from Claude Max subscription
- No API key needed, no per-token billing
- Usage counted as "messages" not "tokens" in Claude Max dashboard

---

**Updated:** 2026-02-17  
**Status:** ✅ Active  
**Cost:** €0 extra (included in subscription)