Claude 101 · lesson 2 of 6
Claude Model Family
The three tiers
| Tier | Tagline | Use for |
|---|---|---|
| Opus | Most capable | Hard reasoning, agents, complex code, long-horizon planning |
| Sonnet | Balanced | Production workloads, RAG, most agent loops, day-to-day API use |
| Haiku | Fastest & cheapest | Classification, routing, high-volume tasks, edge latency-sensitive flows |
Current model IDs (Claude 4 family)
- Opus 4.7 —
claude-opus-4-7 - Sonnet 4.6 —
claude-sonnet-4-6 - Haiku 4.5 —
claude-haiku-4-5-20251001
Always pin the model ID in production. When migrating, run evals on the new model before flipping traffic.
Picking a model
A pragmatic decision tree:
- Is this a high-volume routing/classification step? → Haiku.
- Does it need deep reasoning, tool use across many turns, or long-horizon planning? → Opus.
- Everything else → Sonnet (default for most production agents).
For agent loops the standard pattern is:
- Haiku for cheap subagents and routers.
- Sonnet for the main worker.
- Opus for the orchestrator on hard problems.
Tradeoffs to internalize
- Capability ↑ → cost & latency ↑.
- A bigger model with no prompt caching can be slower and pricier than a smaller model with caching.
- Extended thinking trades latency for accuracy on hard problems — don't enable it for trivial calls.
See also
- messages-api — how to specify model in requests
- prompt-caching — cost levers
- extended-thinking — when to enable thinking