Anthropic offers Claude in tiers: Haiku (small and fast), Sonnet (the balanced middle) and Opus (the largest). Newer versions replace older ones over time, but the tiers stay useful as a mental model.
Haiku: speed and volume
Good for classification, extraction, routing, autocomplete-style tasks, short replies and anything you run thousands of times. Example: claude-haiku-4-5 at $2.10 / $10.50 per 1M tokens (input / output).
Sonnet: the default choice
Strong at coding, writing, analysis and conversation, at a moderate price. Most apps should start here. Example: claude-sonnet-4-6 at $6.30 / $31.50 per 1M tokens (input / output).
Opus: the hardest problems
For complex reasoning, difficult debugging, long careful writing, and tasks where a mistake is expensive. Example: claude-opus-4-7 at $10.50 / $52.50 per 1M tokens (input / output).
Mixing tiers
You rarely need one tier for everything:
- Router + worker: Haiku decides what kind of request it is; Sonnet handles normal ones; Opus only the hard ones.
- Draft + review: Sonnet drafts; Opus reviews only high-stakes output.
- Agents: coding agents like Claude Code let you set a smaller model for quick steps and a larger one for planning (setup).
How to decide
Run your own examples through each tier (method). If the smaller tier is good enough, the savings add up quickly at scale.
Claude, GPT, Gemini, DeepSeek, image and video models through one OpenAI- and Anthropic-compatible endpoint. Prepaid, pay only for what you use.