Many articles describe the difference between Opus and Sonnet as “high capability” versus “balanced,” but that doesn’t help you choose. The useful criterion is the multiplier. When you run the same one-step task on both, the price difference is about 1.7 times. Knowing that figure turns “Which one is smarter?” into a question you can answer: “Is this task worth paying 1.7 times as much for?”
One more point: many of the top search results are based on models from one or two generations ago. The prices on this page are loaded from the catalog when it renders, so the figures change along with any price cuts or new generations.
Current model rates and the actual cost of one step
Per 1M tokens, and for one step of 25,000 input / 1,200 output tokens:
| Model | Input / output (per 1M) | One step | Best for |
|---|---|---|---|
| Claude Haiku 4.5 | $0.70 / $3.50 | About $0.0217 | Summarizing, classifying, formatting, and extracting |
| Claude Sonnet 4.6 | $2.10 / $10.50 | About $0.0651 | Everyday coding in general (set as default) |
| Claude Opus 5 | $3.50 / $17.50 | About $0.1085 | Design decisions, unexplained bugs, and multiple constraints |
Choosing a model: default to Sonnet and switch to Opus when you get stuck
There’s one practical approach: set Sonnet as the default and switch to Opus only when you get stuck. If you do the reverse, you’ll pay the top-tier rate even for tasks with clear steps.
Choose based on the nature of the task, not how difficult it feels.
- Sonnet is enough — The requirements are clear. The steps are known. You’re adding to existing code. The reason a test failed is clear.
- Switch to Opus — You need to make a design decision. The cause hasn’t been identified. You need to meet several constraints at once. Sonnet has failed in the same place twice.
- Switch to Haiku — Summarizing, classifying, formatting, and extracting. The agent’s internal calls.
The cost of switching is almost zero. From the API’s perspective, it’s just a model name string; the endpoint, key, and request format stay the same. So when in doubt, it’s faster to try it.
Choose based on your own repository, not benchmarks
Public benchmark results change with the choice of tasks, and they go out of date faster than models are updated. By the time you read a ranking here, it may already be about a different generation.
Here’s a 30-minute process instead. Pick three problems you’ve actually struggled with, give both models the same instructions, and track whether they solve them, how many revisions are needed, and how many tokens they use. This gives you a ranking for your own use cases and shows whether it’s worth paying 1.7 times as much.
For an overview of model selection and pricing in Claude Code, see Claude Code pricing. For options when you hit a usage limit, see Claude Pro limits. For a detailed English version, see Claude Opus vs Sonnet vs Haiku.
Frequently asked questions
What’s the difference between Claude Opus and Sonnet?
They differ in their positioning and price. Opus is the top-tier model for the hardest problems, while Sonnet balances capability and price for everyday work. The price difference is specific: at Kunavo’s rates per 1M tokens, Claude Opus 5 costs $3.50 / $17.50, while Claude Sonnet 4.6 costs $2.10 / $10.50. For the same one-step task, the difference is about 1.7 times.
Which one should I use day to day?
The practical approach is to set Sonnet as the default and switch to Opus only when you get stuck. If you do the reverse, you’ll pay the top-tier rate even for simple tasks. As a rule of thumb, use Sonnet for tasks with clear specifications and steps, and Opus for tasks that require design decisions, have an unknown cause, or need to satisfy several constraints at once.
When should I use Haiku?
These are routine tasks such as summarizing, classifying, formatting, and extracting. At $0.70 / $3.50 per 1M tokens, the price is on a completely different scale, so moving these auxiliary tasks to Haiku alone noticeably lowers your bill. A setting that routes Claude Code's internal calls to Haiku (ANTHROPIC_DEFAULT_HAIKU_MODEL) is also required when using Kunavo. If it is not set, the haiku alias calls a model ID that Kunavo does not offer (Haiku 5.5 from v2.1.293 onward), causing /model haiku and subagents configured with haiku to return 404s.
What changes when I switch models?
From the API’s perspective, you just change the model name string. The endpoint, key, and request format stay the same. That also makes experimentation inexpensive—if you’re unsure, send the same input to both models once and compare the results and usage. That’s the quickest way to decide.
Can benchmarks tell me which one is smarter?
Public benchmark rankings change with the choice of tasks, and the results have a short shelf life given how often models are updated. A more reliable approach is to test them on your own use cases. Pick three problems you’ve actually struggled with, give both models the same instructions, and track whether each solves them, how many revisions are needed, and how many tokens they use. In about 30 minutes, you’ll have a ranking for your needs.
As of when are the prices in this article?
Prices aren’t hard-coded in the text; they’re loaded from the model catalog when the page renders. That means the figures in this article change along with model additions, price cuts, or new generations. Avoiding outdated figures when a new generation arrives is exactly why this approach was chosen.