## The Hidden Multiplier in Claude Max
Claude Max 20x subscribers get separate weekly quotas for Opus and Sonnet. Most people use one or the other. The real play is using both strategically — and it effectively turns your 20x subscription into something much bigger.
## The Math
API pricing per million tokens:
| Model | Input | Output | Relative Cost |
|-------|-------|--------|---------------|
| Opus | $15 | $75 | 5x |
| Sonnet | $3 | $15 | 1x |
On Max 20x, both models get generous weekly quotas. Sonnet is 5x cheaper on raw API, so the effective value multiplier on Max is:
- **Opus on Max 20x** = 20x value vs API pricing
- **Sonnet on Max 20x** = effectively ~100x value vs API pricing
And they draw from **separate pools**. You're not choosing — you get both.
## The Strategy
### Use Opus for:
- Architecture and planning decisions
- Complex multi-step debugging
- Code that requires deep reasoning about system interactions
- Orchestrating parallel workflows
- Writing nuanced, context-dependent code
### Use Sonnet for:
- Translation and localization (proven at scale)
- Bulk code modifications (find pattern → apply change)
- Repetitive refactoring across many files
- Subagents doing well-defined tasks
- Any work where the instructions are clear and mechanical
### The Orchestrator Pattern
The optimal setup we discovered:
1. **Main session runs on Opus** — plans the work, makes architectural decisions, orchestrates
2. **Subagents run on Sonnet** — execute well-defined tasks in parallel
3. **Results flow back** — Opus merges, validates, and commits
This is like having a senior architect directing a team of capable developers. The architect doesn't need to write every line — they need to make the right decisions and coordinate.
## Real-World Example: 5-Language Translation
We translated 14,340 UI strings across 5 languages (Spanish, German, French, Japanese, Portuguese) in a single session:
- **Opus (main session):** Generated .po files, wrote chunking scripts, created storage objects, designed the merge pipeline, committed and pushed code
- **Sonnet (24 subagents):** Translated 6 chunks per language, stored results via MCP objects, posted progress to the feed
**Result:** 50% of Opus session used, only 5% of Sonnet weekly quota touched. The heavy token work (translation output) came from the cheap pool.
**Equivalent cost on raw API:** $750-$1,500 estimated
**Equivalent human effort:** 10-13 weeks of professional translator time
**Wall clock time:** ~2 hours
## When You're Hitting Limits
If your Opus weekly quota is getting tight:
1. Switch your main session to Sonnet for mechanical tasks
2. Keep launching Sonnet subagents for parallel work
3. Save remaining Opus quota for decisions that actually need it
Think of it as two fuel tanks. When tank A runs low, you've still got a full tank B for the right kind of work.
## Key Insight
The subscription cost is fixed at $200/month. The value you extract depends entirely on how you split work between the two models. Using only Opus leaves the Sonnet quota on the table. Using both strategically maximizes your weekly throughput.
**20x on Opus. 100x on Sonnet. Both included.**