Claude Opus vs Sonnet: Which Model to Use

Use Sonnet 5.5 for everyday, well-scoped work: it costs half as much as Opus 5.5 ($2 and $10 per million tokens against $4 and $20), and it is faster. Use Opus 5.5 for long-running agentic coding, hard debugging and work where you would rather spend more than redo it. Anthropic itself says to start with Opus for most workloads and move down to Sonnet where speed and cost matter. Checked on 31 October 2026.

The short version

The current lineup

Anthropic's models overview lists four current models. Opus and Sonnet are the middle two.

Two small details deserve attention. First, the default effort differs: Opus 5.5 defaults to medium and Sonnet 5.5 to high. That means out of the box Sonnet may think harder on a given prompt than Opus does, so "Opus is smarter" is a statement about ceiling, not about default behavior. Second, thinking on Opus is always on, while on Sonnet it is adaptive and can be tuned.

All current models accept text and image input, produce text output, and support tool use. The knowledge cutoff for all four is June 2026.

What Anthropic says each is for

Opus 5.5

The Opus page calls it "our most capable Opus model yet for coding, agents, and knowledge work." It lists these use cases:

• Agentic coding. Long-running work in large codebases: building features, debugging, refactoring and code review. The page says it finds the root cause before changing anything and checks its work as it goes.
• Agents. Orchestrating complex multi-tool tasks, coordinating subagents and using memory across sessions.
• Enterprise work. Instruction following, staying in scope and producing spreadsheets, slides and docs.
• Finance. Reading dense filings, models and charts.
• Vision and computer use. Reading dense documents, charts and screenshots.

It also says Opus 5.5 costs less per token than Opus 5 and uses fewer tokens per task, so work billed by token costs about 40 percent less than on Opus 5. That is a comparison with the previous Opus, not with Sonnet.

Sonnet 5.5

The Sonnet page calls it "a faster, lower-cost complement to Opus 5.5" that "excels at well-scoped everyday tasks across coding, agents, and knowledge work." It lists:

• Everyday development: building features, fixing bugs, reviewing code, with edits small enough to review.
• Well-defined agent tasks such as investigation, review and drafting.
• Interface design, diagrams and slides that need little editing.
• Knowledge work, analysis and documents.

Sonnet 5.5 is priced at $2 and $10 per million tokens, with up to 90 percent savings from prompt caching and 50 percent from batch processing. Anthropic says it costs up to an estimated 30 percent less to run than Sonnet 5 on typical workloads billed by token.

The wording tells you the intended split. Opus is for work that runs long and has to hold together. Sonnet is for work with a clear scope.

Cost in practice

The API prices are clean: Opus is exactly twice Sonnet on input, output, and also on cache pricing.

Batch processing saves 50 percent on both. US-only inference is available at 1.1x pricing. Fast mode for Opus 5.5 is available in Claude Code and on the Claude Platform at up to 2.5x faster speed, priced at $8 input and $40 output per million tokens, which is double the standard Opus rate.

A model that costs twice as much per token is not always twice as expensive per task. Anthropic's own wording for Opus says it uses fewer tokens per task than the earlier Opus, and for Sonnet that it needs far fewer tokens than the earlier Sonnet on long tasks. Whether Opus finishes a job in fewer tokens than Sonnet is a question about your job, and neither page answers it. We did not measure it. Run a sample of your real tasks on both and compare the bill and the number of retries.

Subscription plans

On claude.ai, the plan decides which models you can reach.

Fable is paid with usage credits on Pro and included up to half your weekly limit on Max and Ultra. All paid plans share one usage pool across chat and Claude Code, and how fast you reach a limit depends on the model you use. Anthropic says so directly on the pricing page: how fast you reach a limit depends on the model you use, how long your chats are and how large your files are. The pricing page does not give a ratio between Opus and Sonnet consumption.

Using them in Claude Code

Claude Code's model docs describe aliases so you do not have to remember version numbers: default, best, opus, sonnet, haiku, fable and opusplan. The opus, sonnet and haiku aliases resolve to the newest version on the Anthropic API, and to an earlier version on some other providers.

opusplan is the most useful alias for this comparison. It uses Opus during plan mode, then switches to Sonnet for execution. The idea is simple: spend the expensive model on thinking through the approach, and the cheaper one on typing out the edits. You can switch models in a session or set a default, depending on your settings.

A reasonable routine:

1. Start a hard task in plan mode on Opus.
2. Review and approve the plan.
3. Let Sonnet execute the steps.
4. If the diff misses the point, go back to Opus for the next plan.

Anthropic's cost page says Claude Code's average cost across enterprise deployments is around $13 per developer per active day, and that choosing the model is one of the ways to manage spend. It also has recommendations on context management and extended thinking settings.

Choosing by task

That table is our reading of Anthropic's positioning statements, not a benchmark result. If one row fails on your work, trust your results over our table.

Benchmarks and what we can say about them

Both model pages show benchmark charts. These are Anthropic's own results, run with settings that Anthropic describes (for example, adaptive thinking at max effort for Opus). They are useful for seeing how a model improved over its predecessor, and less useful for choosing between Opus and Sonnet on your tasks. We do not copy the scores here, and we did not run our own. Treat any ranking on a single leaderboard as one input.

Where Opus 5.5 is stronger

• Long-running, multi-step agent work, per Anthropic's own positioning.
• Orchestrating subagents and using memory across sessions.
• Dense vision and computer-use tasks.
• Hard debugging where the root cause is unclear.

Where Sonnet 5.5 is stronger

• Price: half of Opus on every token type.
• Speed: Anthropic lists latency as fast, against moderate for Opus.
• Free-plan access: Sonnet is the top model you get without paying.
• Scoped daily work, where the edits are small enough to review.
• High default effort for its price.

Moving between them

On the API, change the model ID from claude-sonnet-5-5 to claude-opus-5-5 or back. Both support the same inputs and tool use. Check three things when you switch:

1. Effort settings. Defaults differ, so set effort explicitly if you want the same behavior.
2. Output length. Both list 128K max output.
3. Cost alerts. Doubling the unit price doubles a bill that stays the same size.

Anthropic publishes a migration guide for moving between model versions, and a deprecations page. Retirement dates for all four current models are no sooner than September and October 2027, so there is no urgency to pin a version today.

A short checklist

• Do I have a task that runs for an hour or more? Try Opus.
• Is the task well scoped and checkable? Start with Sonnet.
• Am I paying by token or by plan? Token cost doubles; plan limits drain faster on heavier models.
• Have I set effort deliberately?
• Did I measure retries, not only price per token?

A worked month, and what routing saves

Take a team agent that spends 200M input tokens and 20M output tokens a month, uncached. These are example volumes, used to show the arithmetic with Anthropic's list prices.

The mixed row is the interesting one. Sending only the hard fifth of the work to Opus raises the Sonnet-only bill by 20 percent, from $600 to $720, while the all-Opus bill is $480 higher than that. Whether the hard fifth really needs Opus is the question the figures cannot answer, and it depends on your tasks. Haiku's row is a floor, not a recommendation: Anthropic positions it for high-volume, latency-sensitive tasks, and most coding agents need more than that.

Caching changes the picture. The cache read price is also 5 percent of input on both models, so an agent that resends a long, stable prefix pays a small fraction of the input line above. Batch halves everything for work that can wait. Check both before deciding that Opus is out of budget.

A test that decides the split for you

You do not need a benchmark. You need your own tickets.

1. Collect twenty tasks from the last month and sort them by size: a few trivial, most medium, a few hard.
2. Run each on Sonnet first, with the same instructions file you use every day.
3. Mark a task as passed if you would merge the diff with at most one correction.
4. Re-run only the failures on Opus, and note which of them Opus fixes.
5. Compute cost per passed task for each model, from the token counts in your usage view.
6. Write the rule: tasks that look like the Opus-only ones go to Opus, everything else stays on Sonnet.

If Opus rescues one failure in twenty, the rule is "Sonnet by default, Opus by hand." If it rescues six, move planning to opusplan and review again in a month.

Questions people ask

Is Opus always better than Sonnet? Anthropic describes Opus for long-running work that has to hold together and Sonnet for well-scoped everyday tasks. It does not say Opus wins every task, and the default effort setting differs, so test on your own work.

Can I switch models mid-session in Claude Code? The model docs describe aliases, and opusplan switches between Opus for planning and Sonnet for execution. Check the docs for the command to change model during a session.

Does Opus cost twice as much per task? It costs twice as much per token. Per task depends on how many tokens each model spends, and Anthropic says newer models use fewer tokens than their predecessors.

Our take

Going by Anthropic's documents, the choice is a pricing and positioning ladder: Haiku for volume, Sonnet for everyday work, Opus for long and hard work, Fable above it. Anthropic's own advice is to start with Opus for most workloads and use Fable when Opus at higher effort still falls short. That advice suits API builders who have evals. For most people using Claude Code on a plan, we would reverse the starting point: run Sonnet by default, and bring in Opus through opusplan or for tasks where Sonnet has already failed once.

The reasoning is plain arithmetic. Opus costs twice as much and, on a plan, spends your usage pool faster. If Sonnet gets a scoped task right, you saved half. If it does not, you lose one attempt and switch.

We have not run either model on a controlled task set, so this page makes no claim about relative accuracy. It reports prices, plan access, positioning and defaults that Anthropic publishes, all read on 31 October 2026.

Related

• Claude Code
• Claude vs Claude Code
• ChatGPT vs Claude
• DeepSeek vs Claude for coding
• Compare AI models
• How we compare