Brain, All-Rounder, Hands: What Three Models Reveal About Your Job
Anthropic ships three Claude models: Opus, Sonnet, and Haiku. Most people default to one model for everything and never think about it again.
That default is a blind spot. Not because it wastes money -- it does -- but because it hides something important. The three models are not just a product lineup. They are a map of intelligence itself, broken into tiers. And when you hold that map up against your own work, the picture is uncomfortable.
The trio
Opus is the brain. Sonnet is the all-rounder. Haiku is the hands.
Opus -- deep reasoning, architecture, strategy. The most expensive because thinking is expensive. You use it when the task requires figuring out why and how, not just what.
Sonnet -- the generalist. Coding, drafting, summarization, moderate reasoning. It handles the bulk of everyday work. Not the sharpest, not the cheapest, but self-sufficient.
Haiku -- fast, cheap, no-nonsense execution. File searches, boilerplate, data parsing, tagging. It does not think. It does. Like a runner fetching files -- no complaints, no wasted motion.
The names are a map
Anthropic named their models after literary forms, and the names tell you exactly what each one is for:
Opus -- a major musical or artistic composition. Grand, ambitious, complex. The name signals heavyweight work. You bring in an opus when the task demands depth and scale.
Sonnet -- a fourteen-line poem with formal structure. Balanced, disciplined, complete within its constraints. Not sprawling, not minimal -- just right for the middle ground.
Haiku -- a poem in three lines, 5-7-5 syllables. Stripped to the essence, no wasted words. Maximum output from minimum input. That is exactly how the model operates.
The naming was not accidental. It is a deliberate hierarchy. And hierarchies reveal what the designer values at each tier.
A cultural lens
If you want a mnemonic, think of it through cultural traditions:
Opus recalls the Talmudic intellectual tradition -- questioning everything, analyzing layer by layer, finding meaning in complexity. That relentless depth of inquiry is exactly what Opus brings to architecture and reasoning tasks.
Sonnet resembles the European civic ideal -- well-educated, responsible, competent across domains. Not the most brilliant specialist, but a reliable citizen who keeps the system running.
Haiku maps to modern Japanese work culture -- precise, efficient, no wasted motion. But not the Japan that produced Basho, who compressed entire worldviews into seventeen syllables. That was deep thinking distilled. Modern Japan kept the discipline but lost the depth -- a culture that follows procedures flawlessly without questioning whether they make sense.
And like in the real world, you need all three to build something great. But you do not need them in equal measure. That is where it gets revealing.
The uncomfortable question
The three tiers do not just describe AI capability. They describe the nature of work itself.
Look at what Haiku handles: searching, parsing, extracting, tagging. Tasks with a known shape and a known answer. No judgment required. Now look at your own workweek and ask honestly -- how much of what you do is the same kind of work? Not the same tasks, but the same nature. Searching for the right file. Copying a pattern you have used before. Writing a function that is structurally identical to the last ten you wrote. Closing a ticket whose solution was obvious before you opened it.
Most engineers would say they spend their days solving problems. But the three-tier model suggests a different picture. A large share of what feels like engineering is actually search-and-assemble -- pattern execution dressed up as problem-solving. The real thinking, the Opus-tier work, is the minority. It always was. We just never had a hierarchy that made the tiers visible.
The distortion
Here is what most people miss. They see "three models, three cost tiers" and think it is a menu. Pick what fits your budget. But it is not a menu. It is an X-ray of the knowledge economy.
Anthropic did not decide arbitrarily that thinking should cost more than executing. They discovered it. Training a model to reason deeply costs fundamentally more than training one to pattern-match. The hierarchy reflects the actual cost structure of intelligence -- and that cost structure maps directly onto human work.
Domain knowledge, taste, systems thinking, ownership -- these are Opus-tier human capabilities. They cannot be commoditized because they require genuine reasoning about the problem space. But the work that surrounds them -- the searching, assembling, formatting, and repeating that fills most of a workday -- that is Haiku-tier. Not because Haiku can do your job today, but because that category of work requires no real thought. And work that requires no real thought is the first to be automated away.
The people who default to one model for everything are making the same mistake organizations make with human talent: treating all intelligence as interchangeable. It is not. It never was. The model lineup just makes the tiers visible.
What the map tells you
Stop defaulting to one model. But more importantly, stop defaulting to one understanding of your own work.
The engineers who will matter are the ones operating at Opus tier -- asking why, not just how. The rest is execution, and execution gets cheaper every quarter.