Claude Opus vs Sonnet: is the bigger model worth twice the price?

That is Anthropic's own advice, and our results back it. The same guide lists, among the reasons to step up to Opus, problems where you've tested with Sonnet and it struggled. The guide now talks about the 5.5 generation; the versions we tested are the 4.x ones.

Why twice, when Opus's list price is $25.00 per million output tokens against Sonnet's $15.00, only two thirds more? Because Opus also wrote more. Over the seven tasks it produced 2,097 output tokens to Sonnet's 1,587, and the same prompts were counted as 862 input tokens on Opus against 671 on Sonnet. A higher rate on more tokens compounds.

Here is the prompt: five people, Ana, Ben, Chen, Dev and Eli, each need one 30-minute slot between 9:00 and 11:30. Ana can't go before 10:00, Ben must go right after Chen, Dev must be first or last, and Eli can't have 9:30. Give one schedule, then say how many valid schedules exist.

Both models gave the same valid schedule. Then they counted. Opus split the problem sensibly, by whether Dev goes first or last, and then miscounted both halves: “Dev first (slot 1): Chen-Ben in slots 2-3 or 3-4.” (only slots 2 and 3 work) and ended on “Total: 11 valid schedules.” Sonnet reached “Total: 5 valid schedules. The example above is not the only one.” Its own breakdown was off too (it gave 3 and 2 where the real split is 2 and 3), so it got the total right with a shaky working.

The Jaipur trip was the clearest case. Asked for a 3-day plan for two on ₹15,000, Opus listed every cost line, and its days (₹3,200, ₹3,800 and ₹4,600) add up to its stated total of ₹11,600. Sonnet's plan had a summary table of ₹11,000 and day totals that add up to ₹7,800. One of those is wrong, and nothing in the answer tells you which.

On the memo summary, Opus kept every fact, including the overtime line Sonnet dropped, and didn't relabel anything. Sonnet called the results “Productivity mixed” and changed "answered" to "resolution", small liberties in a task that said to add nothing. On the code fix, both found all three bugs, but Opus raised a clear error on bad input and added “if duplicates should count, drop set() but keep the copy and length guard”, which is the question a careful reviewer would ask.

On the jug puzzle, which has no solution, Opus said no in its first line. Sonnet opened with “Solution: Yes, 4 litres can be measured.” and corrected itself two lines later.

For the firm invoice email, we preferred Sonnet. It was the only one of the two to name a consequence, “If I don't hear back by the end of this week, I'll need to explore further steps to settle the outstanding balance.” Opus was the softest of the five models we tested this week, with “I understand things can get busy” and no deadline at all. Both explained compound interest correctly; Opus was exact to the paisa (Rs 12,597.12) and finished with a rocket emoji, Sonnet rounded to Rs 12,597. Call that a tie.

Each prompt went once to anthropic/claude-opus-4.8 and anthropic/claude-sonnet-4.6 through OpenRouter's API on 7 November 2026, between 16:34 and 16:39 UTC. No system prompt, no extended thinking for either model, answers capped at 700 tokens. Extended thinking might change the puzzle result for either model; we didn't test it. Run a prompt again and the wording, and sometimes the verdict, will change, so read these as one careful afternoon of testing. Quotes are cut where you see [...], and the models' long dashes appear as commas; otherwise they are word for word.

None of this went through AI4Chat. If you want Opus or Sonnet themselves, Anthropic's Claude apps and API are where to get them.

The newer pair keeps roughly the same gap: Opus 5.5 lists at twice Sonnet 5.5's per-token price. Both are cheaper than the 4.x versions we tested, and there is also Haiku 5.5 at $0.10 and $0.50 for quick jobs. We haven't run these prompts through any 5.x model yet. If you're weighing Claude against other families entirely, see ChatGPT vs Claude and Claude vs Gemini.