Choosing a Claude Code model: Opus, Sonnet, or Haiku

The lineup, with real numbers

• Output is priced five times input on every model in the table, so the output column is where the difference actually bites.
• Cache reads bill at 10 percent of the base input rate, which is why a long cached session is far cheaper than the input column alone suggests.
• The Batch API halves both input and output, but it is asynchronous and irrelevant to interactive coding.
• Haiku 4.5 does not support effort levels. Opus 5, Sonnet 5, and Fable 5 all do, and default to high.

Aliases, and why yours may not mean what you think

Claude Code takes either a full model name or an alias. Aliases track the recommended version and change over time, which is convenient until you are on a provider where they resolve somewhere else.

Setting and switching it

# 1. in session, effective from the next turn
/model sonnet
/model claude-opus-5
/model               # opens the picker; Enter saves as default, s is session-only

# 2. at launch, this session only
claude --model opus

# 3. environment, this session only
ANTHROPIC_MODEL=claude-haiku-4-5 claude

# 4. settings file, persistent
#   { "model": "claude-sonnet-5" }

• /model saves your choice as the default for new sessions by writing the model field in user settings. Press s in the picker to switch for this session only.
• The picker asks for confirmation when the conversation already has output, because the next response re-reads the full history without cached context. That confirmation is a real cost warning, not a formality.
• Resumed sessions keep the model they were using when the transcript was saved, so another terminal switching models cannot change yours on resume.
• To run different models in different terminals at once, launch each with its own --model flag rather than switching with /model.
• A subagent can pin its own model in frontmatter, and a skill can override the model for the turn it runs in.

A routing table

What the gap costs on one real task

List rates per million tokens are hard to reason about, because an agent session is mostly cache reads. Here is the same task priced four ways, so the ratio is concrete rather than notional.

The task: 300,000 input tokens across the session, of which 90 percent arrive as cache reads, plus 20,000 output tokens. That is a realistic shape for a bounded feature or a substantial debugging run.

On a subscription the same ratios apply to your usage window rather than to a dollar figure: a session or weekly allowance is consumed roughly in proportion to token cost, and the windows are shared across models. Switching to Sonnet with /model after hitting an Opus-specific limit keeps you working; it does not restore a session or weekly window, because those are not per model.

Escalate on evidence, not on anxiety

Anthropic publishes two pieces of guidance that read differently, and both are correct in context. The platform models page says to start with Opus 5 for complex agentic coding and enterprise work. The Claude Code cost guidance says Sonnet handles most coding tasks and to reserve Opus for architectural decisions and multi-step reasoning. The first is answering "which model is capable of this"; the second is answering "which model should run your Tuesday".

The efficient pattern resolves both. It is not choosing a model per project, it is starting cheap and escalating the moment a task has shown you it is hard.

1. Start on Sonnet 5. It handles the large majority of work at a fifth of Opus output price.
2. Escalate to Opus 5 when Sonnet has produced a wrong answer twice, when the task is architectural, or when you cannot describe the fix yourself.
3. Drop to Haiku 4.5 the moment a task turns mechanical, which frequently happens partway through: Opus decides the approach, Haiku applies it across thirty files.
4. Escalate for review even when the work was done cheaply. A fresh Opus read of a finished diff in a subagent is one of the best uses of the expensive model there is.
5. Or let opusplan do it. It runs Opus during plan mode and Sonnet for execution, which is the same discipline with no manual switching.

Model choice also interacts with reasoning effort, and the two multiply. High effort on Opus 5 is many times the cost of the default on Sonnet 5 for work the second combination often completes the same way. Decide the model first, then the effort, and change one at a time.

Questions people ask

Keep reading