You need to get from A to B. You can take the bus, or you can take a sports car. Both arrive at the same place. One costs $2.50, the other costs $200 in fuel.
The sports car has 600 horsepower, heated leather, and a turbo that kicks in at 4,000 RPM. None of that matters if the destination is the grocery store.
That's the difference between running a task with Opus and running it with a Haiku workflow.
The multiplier
Anthropic publishes per-token pricing for each model. Normalised against Haiku as a baseline:
- Haiku: 1.0x
- Sonnet: 3.75x
- Opus: 18.75x
That's the floor. If Opus and Haiku used the exact same number of tokens on the same task, Opus would cost 18.75 times more. Same destination, same route, 18.75x the fuel bill.
Toggle between "equal tokens" and "realistic tokens" above. Equal tokens shows the pure rate difference. Realistic tokens shows what actually happens.
The toil tax
Here's what "realistic" means: Opus doesn't use the same number of tokens.
Higher-capability models try harder. That's the point. They reason deeper, explore alternatives, self-correct. On ambiguous problems, this is why you pay for Opus. On well-defined tasks, it's waste.
Give Opus "scrape these 5 URLs and extract the pricing table from each." It'll do the job. But it'll also consider edge cases you didn't ask about, retry with alternative selectors when the first one worked fine, and narrate its reasoning along the way. More tokens, same result.
A Haiku agent with the same instructions follows the instructions. Fewer tokens. Same output.
The sports car doesn't just burn premium fuel. It takes scenic detours.
Workflows are the bus route
A bus doesn't improvise. It follows a fixed route, makes its stops, and gets there. That's a workflow.
When you break a task into three focused steps, each with a specific agent, specific tools, and a clear prompt, you've already done the hard reasoning. The agents don't need to figure out what to do. They need to follow directions.
- Step 1: Haiku agent scrapes URLs, writes results to the shared volume
- Step 2: Haiku agent reads those files, extracts data points
- Step 3: Haiku agent formats the final report
Each step is simple enough that Haiku handles it without breaking a sweat. The workflow is the reasoning. The agents are execution.
Three Haiku steps at 1.0x each, versus one Opus agent at 18.75x burning extra tokens on exploration it was never asked to do.
You don't have to draw the route
The obvious objection: "Setting up three agents is more work than running one." Fair. That's why every create button in Svortie has a "Create with AI" option.
Describe what you want done. The platform makes one Sonnet call to decompose your description into focused steps, creates the agents, wires the workflow, and runs it. All Haiku. One smart call to plan the route, then cheap specialists execute each stop.
You get the cost structure of the bus without having to be the one drawing the map.
When Opus earns its cost
The sports car isn't always wrong. Sometimes you're on the autobahn.
Opus earns its cost when the task is genuinely ambiguous and you can't decompose it upfront. Complex code review where context spans thousands of lines. Analysis where the "right approach" depends on what you find along the way. One-shot problems where the reasoning IS the work.
But those tasks are rarer than most people think. If you can describe the steps, you can build a workflow. And if you can build a workflow, you don't need Opus.
Every agent in Svortie has a model picker. You can set any step to Opus, Sonnet, or Haiku. The option is there. If you want to pay 18.75x more for the same result, it's your credits.
Decomposition is cheaper than reasoning
That's the philosophy. Not "use cheaper models." Not "spend less." The claim is stronger: breaking complex work into simple steps produces better results at a fraction of the cost.
A single expensive agent has to hold the entire problem in its head, planning, executing, and course-correcting all at once. A workflow separates the thinking from the doing. The platform handles decomposition. Cheap specialists handle execution. State passes between steps via the filesystem, not the context window.
Complex work, simple agents, small bill. That's the point.