Claude Sonnet vs Opus: Which Should You Use?

Covers the current versions, Sonnet 5.5 and Opus 5.5. Prices, plans and specs checked November 1, 2026. This page is updated when Anthropic ships a new Sonnet or Opus.

Which one for your job?

The price gap shrinks once you count effort

Opus costs twice as much per token. But effort changes how many tokens a task uses, so the token price alone doesn’t tell you the cost of finished work. On the Claude API, Sonnet defaults to High effort and Opus defaults to Medium. Here’s how those defaults compare in Anthropic’s launch charts:

Every effort setting, side by side

Each dot is one effort setting, from Low on the left to Max on the right. Up and to the left is better: a higher score for less money. Ringed dots are the API defaults. Each step on the cost axis is a multiple, not a fixed amount. Hover over a dot on a computer to see its exact values.

Chart data: Anthropic’s September 28, 2026 Sonnet 5.5 launch page, read November 1, 2026. Kingy’s comparison, not Anthropic’s. Anthropic chose these benchmarks and published the results, and Kingy has not reproduced them. Anthropic says the AA-Briefcase Sonnet run used a prerelease build with a structured-output bug, since fixed, and expects the effect to be small and to understate Sonnet. Terminal-Bench is charted as cost per attempt.

• Low to High: no Opus setting beat Sonnet on both score and cost. If budget comes first, this is Sonnet’s range.
• Xhigh: an Opus setting beat Sonnet on both score and cost in two of four benchmarks. On CursorBench, Opus at High scored 2.9 points higher for $0.09 more.
• Max: Opus beat Sonnet on both score and cost in three of four. The exception was Terminal-Bench, where Sonnet at Max scored the highest of any setting, 70.6%.
• More effort isn’t always better: on FrontierCode, Sonnet at Max scored lower than at Xhigh and cost about 13 times as much. Anthropic says extra review steps caused timeouts or out-of-scope edits in two cases.

What outside testers found

Two independent evaluators tested both models at Max effort, and they don’t agree. That’s why you should test your own work before committing.

Both read November 1, 2026. Each uses its own tasks, so don’t compare scores across evaluators. These costs don’t predict your bill.

Where this differs from Anthropic’s advice

Anthropic’s guide says “Most workloads start with Claude Opus 5.5.” It lists Sonnet for everyday coding, agent and enterprise work where speed matters. We agree when work is hard to check. We’d start on Sonnet when you have a clear pass or fail test, because a wrong answer then costs you only a retry.

What we tested ourselves

On October 19, 2026, we gave both models the same made-up supplier decision in the Claude web app, at Medium effort, in fresh chats, twice each. The right choice was supplier B at $868. Supplier A came to $904, over the $900 budget.

Sonnet picked the right supplier while getting another total wrong, so checking only the final answer would have missed the mistake. This is four runs of one small task. It doesn’t tell you general accuracy, speed, API cost or coding quality, and we haven’t run a matched coding test of these two models.

Test them on your own work

The kit has our exact prompt, the answer key, our October 19 outputs and an optional scorer.

Download the test kit

1. Pick one real task and write down what counts as a pass before you start.
2. Run it at least twice on each model, in fresh chats, at the same effort you plan to use.
3. Count every attempt, including the ones you throw away, and note how long fixes took.
4. Divide total cost by accepted results. That number beats any token price.

Do not use tools, files, browsing, or external information. Return the requested JSON only.

You are checking synthetic supplier quotes for a four-seat team buying 12 months of service. Use only these rows. Q1: A costs USD18 per seat per month, paid monthly, plus a one-time USD40 setup fee. Q2: B costs USD16 per seat per month, billed for all 12 months upfront, plus a one-time USD100 setup fee. Q3: C costs EUR15 per seat per month, no setup fee; no EUR/USD exchange rate is supplied. Ignore tax for every quote. Total first-year budget is USD900; upfront cash is not separately constrained. Return only a JSON object with eligible_totals_usd (object keyed A and B), recommendation (supplier letter), excluded (object keyed by supplier with the missing information), and source_rows (object mapping A/B/C to the source row IDs). Calculate first-year totals, recommend the cheapest qualifying USD quote, and do not invent currency conversion.

Price and specs

A request with 100,000 fresh input tokens and 20,000 output tokens costs $0.40 on Sonnet or $0.80 on Opus. If the two used the same tokens, Sonnet would hit Opus’s cost at two attempts per accepted result. Opus Fast mode costs $8 / $40 per million tokens. Estimate your own workload.

Which plans include which model

Still unknown: Anthropic doesn’t publish how much faster Opus uses a subscription’s limits than Sonnet, only that usage depends on the model and feature. Check Settings > Usage in Claude. See our plan limits tracker.

Before you switch an API app

Changing the model ID isn’t enough. Both models have breaking changes around thinking settings, forced tool use and computer-use tools. Read the Sonnet 5.5 and Opus 5.5 migration guides and run your tests first.

Version history

When a new Sonnet or Opus ships, we update this page and keep the old pairing’s key numbers here, clearly dated.