Opus, Sonnet and Haiku aren't three separate AI assistants. They're one and the same Anthropic model, made in three sizes - like one dish served in three portions at three different prices. You choose between them exactly like you'd choose a portion: speed, price, and how deeply the model thinks.
Picture one dish the chef offers in three portions. A kid's portion, a regular one, and the tasting-menu portion with everything on top. The ingredients are the same. The recipe too. The difference is how much labor, attention and raw material goes into the plate, and, accordingly, how much you pay. That's exactly how a model family like Claude works. Opus, Sonnet and Haiku go through the same training and see the world in a similar way, but with different depths of thinking and different speeds.
In practice, choosing between the three is a decision the programmer makes while building a product. Building a chatbot that answers ten thousand customers a day with simple questions like 'when do you open', you use Haiku. It's fast, it's cheap, and it doesn't need more than that. Analyzing a contract, writing complex code, or explaining a financial report, you use Opus, because there the mistake is costly and it's worth paying for a deeper model. For everything in between, writing text, answering emails, ordinary coding, it's Sonnet. That's the model most people actually meet when they open Claude, without ever thinking about which model they're using.
The reason three sizes exist is simple math. Not every task deserves the most powerful and most expensive tool. Opus costs about five times more than Haiku for the same amount of text, because it also carries about five times more computing power behind it. You pay for the power you actually use, not for the brand.
Two different audiences use them. Ordinary people who just open Claude in the browser or the app usually get Sonnet by default, and Opus kicks in when the task gets heavy. Companies and programmers building products on top of Claude pick the model by hand for every function in their system, exactly the way I pick a different model for a different task when I'm building software.
How I read it
One of the things I learned building systems with these models: size isn't a marketing gimmick, it's an engineering decision. Put Haiku where Opus should be, and the system makes dumb mistakes. Put Opus everywhere, and your product gets slow and expensive for no reason. Knowing how to pick the right size for the right task counts for almost as much as the model itself.
So when you hear someone say 'I use Claude', the question worth asking is which Claude. The answer says a lot about what they're actually doing with it.