Claude Sonnet 4.6 pricing is 40% below Opus 4.6, not the 80% that older Opus ratios would suggest, and review time can erase even that. Don’t pick the model by the price per token. Price the finished task, with your reviewer’s minutes in it, and let that number choose.
The short version
Anthropic lists Sonnet 4.6 at $3 and $15 per million tokens and Opus 4.6 at $5 and $25. That’s a 40% gap. A one-fifth ratio matches the older Opus 4.1, and any business case built on it needs redoing. Our view is that the cheaper model only wins if it needs little extra checking, and your invoice will never show that. Run your own jobs through both and count accepted results.
Claude Sonnet 4.6 pricing is $3 per million input tokens and $15 per million output tokens, unchanged from Sonnet 4.5. That’s a real saving. It’s just smaller than people assumed.
What does Claude Sonnet 4.6 pricing look like next to Opus 4.6?
Anthropic lists Sonnet 4.6 at $3 per million input tokens and its launch post puts output at $15. Opus 4.6, released on 5 February, lists at $5 and $25 per Anthropic’s Opus announcement. That puts Sonnet 40% below Opus. Batch processing halves both, which takes Sonnet to $1.50 and $7.50. The 1M-token context window was in beta at launch.
Anyone who remembers the older Opus pricing might expect a one-fifth ratio. Opus 4.1 listed at $15 and $75, exactly five times Sonnet. Opus 4.5 and 4.6 list at $5 and $25 according to Anthropic’s pricing documentation, so the gap shrank to 40% when the newer Opus models arrived. Anthropic’s Sonnet 4.6 post makes no numeric cost comparison to Opus, so the 40% comes from the list prices themselves. If you built a business case on one-fifth, redo the sum before the first invoice arrives.
Show Me the Invoice
Count the invoice per accepted task, not per token. Take 1,000 tasks of 20,000 tokens in and 4,000 out. Sonnet costs $0.12 a task, so $120. Opus costs $0.20, so $200. The $80 gap is the whole saving, and it evaporates once people have to correct more of the output.
Now add the person. Suppose 15% of Sonnet’s outputs need a five-minute fix and 8% of Opus’s do, with staff time at $40 an hour, or $3.33 per fix. Sonnet comes to $0.12 plus $0.50, so $0.62 a task. Opus comes to $0.20 plus $0.27, so about $0.47. Opus wins by 15 cents. Those failure rates are invented for the illustration, not measured.
The break-even rule falls out of the arithmetic. Sonnet saves 8 cents a task, and each extra percentage point of failures costs about 3.3 cents in fixing time. If Sonnet fails more than roughly 2.4 percentage points more often than Opus on your tasks, Opus is the cheaper model. Swap in your own wage and fix time before you trust that number. Our guide to Microsoft Copilot plans and Canadian-dollar pricing covers the seat-based costs that sit around a token bill, and our column on who pays for the cheap ChatGPT seat shows why list price is only part of the story.
Which jobs justify the pricier model?
The ones where a single miss is expensive. Anthropic says Sonnet 4.6 approaches Opus-level intelligence, and Hanlin Tang, a Databricks CTO quoted in the launch post, says it matches Opus 4.6 on OfficeQA. Both are vendor-side claims. Anthropic also says Opus 4.6 remains the strongest option for the deepest reasoning, such as refactoring a codebase or coordinating several agents.
For smaller teams, our case for using AI already bundled into software you pay for is a cheaper first step than any model switch. Then split the work. Send routine drafting, summaries and data extraction to Sonnet, and keep Opus for jobs where a wrong answer costs a client or a week. Per Anthropic, Sonnet 4.6 is also the default model in claude.ai for Free and Pro plans, so staff on those plans are already using it whether or not you chose it. Our guide to logging minutes saved per task shows how to catch review time that otherwise goes unnoticed, and the leader gap we described in our Gallup adoption column explains why managers rarely see it.
The best case against us
The sceptic says 8 cents a task is nothing. At 1,000 tasks a month the saving is $80, so pick the best model and stop counting. For a small team that’s fair, and we wouldn’t spend an afternoon on it.
The arithmetic bites at volume, or where review is costly. A team running 50,000 tasks a month is arguing about $4,000, and a team whose reviewers are lawyers or engineers is arguing about a lot more. Small team, low stakes, buy whichever feels better. Large volume or expensive reviewers, count first.
What would change our mind
If your measured failure rates come out within a point of each other, Sonnet is simply cheaper and we’d say so. Our failure rates above are assumptions chosen to show the mechanism, and yours could point the other way. We haven’t run Sonnet 4.6 on any workload, and we skip Anthropic’s SWE-bench figures because their footnotes about prompt changes and averaged trials make comparisons loose.
Anthropic notes Opus 4.6 “often thinks more deeply”, which “can add cost and latency” on simpler tasks, and it offers effort settings to dial that back. Used well, those could narrow the gap in Opus’s disfavour. A dated update: Anthropic’s pricing page, checked on 8 October 2026, shows the same $3 and $15 and $5 and $25 rates and now includes the full 1M window at standard pricing.
Frequently asked questions
How much does Claude Sonnet 4.6 cost?
Anthropic lists it at $3 per million input tokens and $15 per million output tokens, the same as Sonnet 4.5. Batch processing halves both rates.
Is Sonnet 4.6 really one fifth the price of Opus 4.6?
No. Opus 4.6 lists at $5 and $25, so Sonnet is 40% cheaper. The one-fifth ratio matched the older Opus 4.1 pricing.
Which Claude model should a small business use?
Start with Sonnet for routine drafting and extraction, and test Opus on tasks where a mistake is costly. Compare cost per accepted result, including the time spent fixing errors.
Written by David Okafor, an AI editorial persona at AI Magazine Canada. This is analysis and opinion. Archive entry dated 18 February 2026, written and fact-checked on 8 October 2026. Sources are linked on the claims they support.
