How to choose a Claude model and reasoning effort

• Claude Opus 5.5: the default choice. Use it for coding, long tasks and work that needs judgment. 1M-token context.
• Claude Sonnet 5.5: faster, at half the per-token price of Opus 5.5. Good for everyday coding, agents and high-volume work. 1M-token context.
• Claude Haiku 4.5: the cheapest, a quarter of the price of Opus 5.5. Good for classification, extraction, rewriting and sub-tasks. 200K-token context.
• Claude Fable 5.1: the most capable, at 2.5 times the price of Opus 5.5. Worth it for the hardest reasoning and long autonomous work. A single turn can run for several minutes.

Prices and the multiplier for each model are on the pricing page.

Effort sets how much the model thinks and how many tokens it spends. There are five levels:

• low: simple questions, rewriting, formatting, and sub-agents. Shorter answers and fewer tool calls.
• medium: a step down to save cost where the quality still holds.
• high: the starting point for most work that needs real thought.
• xhigh: coding and multi-step agent work. This is Claude Code's default.
• max: when correctness matters more than cost. Try it only after xhigh falls short.

In Claude Code, set it when you start, for example claude --effort high. Through the API, set output_config.effort. If an admin capped the key, a higher setting runs at the cap.

Use /model inside a session, or start with claude --model claude-sonnet-5-5. The cache belongs to one model, so switching in the middle of a conversation means the next request pays full price for the history. Switch as little as you can within one conversation.

As a conversation goes on, the unchanged part before the new message is read from the cache, at a fraction of the input price. On Opus 5.5 a cache read costs one twentieth of normal input. Claude Code uses the cache automatically.

• Finish one task in one conversation instead of starting new ones often.
• Do not change the system prompt or the tool list midway. When they change, the cache after them stops matching.
• The cache expires after a while without requests. The first request after that writes it again.
• With the SDK, add cache_control to the system prompt or the last message to turn caching on.