Opus or Sonnet in Claude Code: what it costs

How much do Claude Opus 5.5 and Sonnet 5.5 cost?

Claude API list prices per million tokens, checked November 7, 2026. Claude Code bills by tokens on API plans, and on seat plans the same rates show what usage is worth.

Haiku 5.5's rates apply up to 100,000 prompt tokens; longer prompts cost $0.50 and $2.50. Thinking tokens are billed as output, and Opus 5.5, Sonnet 5.5 and Haiku 5.5 always use extended thinking, so the effort level also moves cost. Sources: Claude models overview and Ganivra's Claude Opus and Claude Sonnet price pages.

What does one Claude Code task cost on Opus vs Sonnet?

Take an illustrative 30-request task: a small change with a few file reads, edits and test runs, the same task we broke down in what Claude Code's starting context costs. Most of its tokens are cache reads, because every request re-reads the conversation so far.

These are illustrative numbers, not a benchmark. Real sessions on the two models won't use identical tokens: one model may take fewer turns or think longer than the other. That is why the comparison that matters is cost per finished task, measured on your own work.

What does the model choice cost a team?

The $0.52 gap per task looks small. Across an illustrative team of 50 developers running 20 such tasks a working day, that is 21,000 tasks a month, and the gap becomes $10,941 a month between an all-Opus and an all-Sonnet team.

Put the other way: every $1,000 of Claude Code usage on Opus would be $500 on Sonnet for the same tokens. On Claude Team seats, the effect shows up as usage limits rather than dollars, because the same work uses up more of each seat's allowance on Opus, so developers reach their limits or usage credits sooner.

When is Opus worth the extra cost?

Compare the token gap with developer time. At an illustrative $100 an hour for an engineer, the $0.52 Opus costs extra on the task above equals about 19 seconds of their time. Opus pays for itself on a task if it saves one retry or a few minutes of the developer fixing the result.

Anthropic gives both sides: its model overview suggests starting with Opus 5.5 for most API workloads, while its Claude Code cost guide says Sonnet handles most coding tasks well. For a team paying per token, a Sonnet default with Opus on demand is the usual compromise.

What does this look like in practice?

Three illustrative patterns that show up when a team looks at Claude Code spend by model:

• Opus left as everyone's default. Most spend sits on Opus, spread evenly across routine work. Switching the team default to Sonnet roughly halves the cost of the same tokens; developers pick Opus for the hard tasks.
• One team needs Opus, the rest don't. A platform team doing migrations gets real value from Opus while product teams writing features and tests don't. Keep different defaults per team in managed settings, and compare each team's cost per developer after a month.
• Subagents inherit an expensive model. A developer switches to Opus, and every subagent that inherits the session model runs on Opus too. Setting model: haiku on simple search subagents cuts that part of the bill to a fraction.

How do I set the default Claude Code model for my team?

Put a model value in managed settings. Claude Code accepts aliases such as opus, sonnet, haiku and opusplan, which follow the latest version for your provider. Deploy it the same way as your telemetry settings: see Claude Code managed settings for cost tracking.

Sonnet by default, Opus available:

{
  "model": "sonnet"
}

Plan with Opus, execute with Sonnet: opusplan uses Opus in plan mode, then switches to Sonnet to carry out the plan.

{
  "model": "opusplan"
}

No Opus at all: availableModels limits which models developers can choose, and enforceAvailableModels (Claude Code 2.1.175 or later) applies the list to the Default option too.

{
  "model": "sonnet",
  "availableModels": ["sonnet", "haiku"],
  "enforceAvailableModels": true
}

A managed model on its own is a starting default, not a lock: developers can still switch with /model, and the managed value applies again at the next launch. Only availableModels restricts the choice. Sonnet 5.5 needs Claude Code 2.1.284 or later and Opus 5.5 needs 2.1.280. Full details: Claude Code model configuration.

What else lowers model cost in Claude Code?

• Put cheap models on subagents. Add model: haiku to a subagent's definition for simple work:

  ---
  name: test-finder
  description: Finds the tests that cover a file
  model: haiku
  ---

• Lower the effort level for simple tasks with /effort. Thinking is billed as output, and on the 5.5 models it can't be turned off, only turned down.
• Clear between unrelated tasks with /clear, so each request doesn't re-read an old conversation.

How do I measure Opus vs Sonnet cost on my team's real usage?

Send Claude Code's OpenTelemetry export to Ganivra and open the Coding agents tab:

• Group by model to see how much of your spend is on Opus, Sonnet and Haiku, and how many sessions and active days each covers.
• Group by developer or team to find who runs on Opus and whether it matches the work they do.
• Read the re-pricing scenarios. Ganivra takes your actual tokens, including cache reads and writes, and prices them on Claude Sonnet 5.5, Claude Haiku 5.5, GPT-5.5 and open models such as GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. That shows what a default-model change would save before you roll it out.

If your product also calls Claude through the API, the Model comparison tab does the same for production traffic: it re-prices each workflow's recorded tokens on current GPT, Claude and Gemini models, and marks where your current model costs more. After a default-model change, compare cost per developer and per session for two to four weeks against the weeks before, and check with the team that work isn't taking longer.

Deciding between seats and API billing too? See Claude Code Team seats vs API.