bereniketech/model-selection

Model Selection

---

1. Model Profiles

| Model | Speed | Cost | Best for | |-------|-------|------|----------| | Haiku 4.5 | Fastest | Cheapest (~20× cheaper than Opus) | High-volume reads, scraping, summarization, classification, simple transforms | | Sonnet 4.6 | Fast | Mid | Implementation, code generation, debugging, most everyday tasks | | Opus 4.7 | Slowest | Most expensive | Architecture decisions, hard debugging, complex reasoning, final review |

Rule: Default to Sonnet. Downgrade to Haiku for volume, upgrade to Opus only when Sonnet fails after 2 attempts or the decision affects the whole system.

---

2. Decision Tree

Task involves architecture / system-wide design decisions?
└── Yes → Opus

Task involves complex debugging that Sonnet can't crack after 2 tries?
└── Yes → Opus

Task involves writing or reviewing implementation code?
└── Yes → Sonnet

Task involves reading many documents / large data to extract a summary?
└── Yes → Haiku

Task involves classification, tagging, or simple extraction at scale?
└── Yes → Haiku

Not sure?
└── → Sonnet

---

3. Sub-Agent Cost Pattern

When orchestrating multi-agent workflows, assign models by role:

Main thread (Opus/Sonnet) — orchestrates, makes decisions, reviews final output
    ↓ spawns
Sub-agents (Haiku) — scrape, read, summarize, classify, filter
    ↓ report back
Main thread — synthesizes Haiku summaries, produces final answer

Example: Scraping 50 articles to find the 3 most relevant: Haiku reads all 50 and returns a ranked list of 5. Sonnet reads those 5 and writes the final synthesis. Cost reduction: ~85%.

---

4. Setting Model per Agent

In Claude Code, specify the model when spawning a sub-agent:

Agent({
  model: "haiku",        // for cheap volume work
  description: "Scrape and summarize 50 articles",
  prompt: "..."
})

Available model values: "haiku", "sonnet", "opus"

---

5. When to Upgrade Mid-Task

Upgrade from Sonnet → Opus when:

• Sonnet has given the same wrong answer 2+ times
• The task is architectural (touching interfaces, data models, module boundaries)
• The task requires cross-file reasoning across >10 files simultaneously
• A mistake here would require significant rework

Rule: Don't use Opus as a default just because it's "better." The quality gap on routine tasks is small; the cost gap is large.

---

6. Cost Reference (approximate)

| Task | Haiku | Sonnet | Opus | |------|-------|--------|------| | Read + summarize 10K tokens | ~$0.001 | ~$0.005 | ~$0.02 | | Write 500-line feature | ~$0.01 | ~$0.05 | ~$0.20 | | Full codebase review (100K tokens) | ~$0.01 | ~$0.05 | ~$0.20 |

Running 10 Haiku sub-agents reading 10K tokens each ≈ cost of 1 Sonnet reading the same.