- In short
- The unified decision logic combines the three task profiles into a single rule: route routine, structured, high-volume tasks to Haiku; route most drafting, synthesis, and analysis to Sonnet; and route complex judgment, high-stakes, or highly ambiguous tasks to Opus. The choice follows the task's profile -- its structure, complexity, and stakes -- not document length, habit, or whichever tier was used last time.
Three profiles, one rule
The three task profiles are only useful together. The CCAO-F exam tests, at the apply level, whether you can fold Haiku, Sonnet, and Opus into a single decision you can run against any task in seconds. That combined rule is the working tool of model selection, and everything downstream in the domain leans on it.
The rule is short: structured, high-volume work goes to Haiku; general drafting, synthesis, and analysis go to Sonnet; complex, high-stakes, or highly ambiguous work goes to Opus. What makes it reliable is not the three branches but the input you feed them -- the task's actual profile. Read the task's structure, complexity, and stakes first, and the branch is determined. Feed it habit or document length instead, and the rule produces the wrong tier.
- Unified model tier decision logic
- A single rule combining the three task profiles: route routine, structured, high-volume tasks to Haiku; route most drafting, synthesis, and analysis to Sonnet; and route complex judgment, high-stakes, or highly ambiguous tasks to Opus. The decision follows the task's profile -- structure, complexity, and stakes -- not document length, habit, or the last tier used.
Reading the profile before the branch
The discipline in applying the rule is to describe the task honestly before choosing. Three questions do most of the work. Is the task structured and run at volume, with low per-item stakes? That is the Haiku branch. Is it ordinary professional work -- drafting, synthesis, analysis -- without extreme ambiguity? That is the Sonnet branch, and it is where most work lands. Does it involve genuinely complex judgment, high stakes, or heavy ambiguity? That is the Opus branch.
Because most professional work is ordinary, Sonnet is the busiest branch and the sensible default, with Haiku and Opus as the deliberate exceptions at either end. The rule is not "pick your favourite tier" or "pick the most capable tier that fits the budget"; it is "match the branch to the profile you just described." When the profile is read accurately, the branch is rarely in doubt.
A useful way to run the rule in practice is directional: start from Sonnet and adjust only when the profile pushes you off it. If quality is falling short on genuinely complex or high-stakes work, move up to Opus. If speed and volume are the dominant requirement and the task is well structured, move down to Haiku. If neither pull is present, you stay on Sonnet. This "Sonnet first, then adjust in one direction" framing keeps the middle tier as the anchor and forces the ends to be earned by a specific profile signal rather than reached for by habit.
What must not drive the choice
The rule's power comes as much from what it excludes as from what it includes. Two inputs are explicitly not allowed to decide the tier. The first is document length. A long document is not automatically Opus work, and a short one is not automatically Haiku work; length says nothing about the structure, complexity, or stakes that the rule actually depends on. Picking a tier by length is a classic misapplication.
The second is habit -- reusing whichever tier you used last time. Yesterday's task and today's may share a topic but differ entirely in profile, and applying the old choice without re-evaluating produces a mismatch. Each task earns its own read. The rule is stable, but its inputs change task to task, so the answer changes too. These two exclusions connect directly to the failure modes in avoiding over-engineering and under-resourcing.
What the CCAO-F exam trips candidates on
Two misapplications are tested. The first is picking a tier based on document length rather than task complexity and stakes. A long routine report is Sonnet or even Haiku work; a short ambiguous judgment is Opus work. Length is a decoy the rule ignores.
The second is applying yesterday's tier choice to a new task without re-evaluating its profile. The tell is a scenario where the task has shifted -- from structured to ambiguous, or from routine to high-stakes -- but the person keeps the previous tier out of habit. The credited answer re-reads the new task's profile and routes accordingly, even when the old choice would have been convenient.
Worked example
Yesterday you used Sonnet for a fifteen-page routine market update and it worked well. Today you face two tasks: a two-page but genuinely ambiguous strategic recommendation with major consequences, and a fifty-page but entirely routine compliance log reformatting job across hundreds of entries. A colleague says 'use Sonnet for both, like yesterday.' How does the rule decide?
The colleague is applying yesterday's choice by habit and, implicitly, leaning on document length -- two inputs the rule forbids. Run each task through the rule on its own profile instead.
The two-page strategic recommendation is short, but shortness is irrelevant. Its profile is genuinely ambiguous with major consequences and no clear procedure -- the Opus branch. Keeping it on Sonnet out of habit would under-resource a task whose stakes and ambiguity call for more depth. The fifty-page compliance log looks heavy, but its profile is routine, structured reformatting repeated across hundreds of entries with low per-item stakes -- the Haiku branch, where speed compounds across the volume. Running it on Sonnet, or worse Opus, would be slower than needed with no quality gain.
So the rule sends the two tasks to opposite ends of the spectrum: the short one up to Opus, the long one down to Haiku. Neither goes to Sonnet, and yesterday's Sonnet choice, made for a differently-profiled task, is irrelevant to both. That is the rule working as intended -- profile in, branch out, with length and habit ignored.
Common misreadings to avoid
Misconception
Longer or more voluminous tasks need a higher tier.
What's actually true
Misconception
Reusing the tier that worked last time is a reasonable shortcut.
What's actually true
How this shows up on the exam
Questions present one or more tasks and ask for the tier, often with length or habit decoys. Describe each task's profile first -- structure, complexity, stakes -- then route: structured volume to Haiku, ordinary work to Sonnet, complex or high-stakes to Opus. The reliable answer re-reads every task fresh and ignores how long the document is or what was used yesterday.
This knowledge point combines the three profiles and unlocks model availability is plan-dependent and the speed-versus-capability tradeoff. It pairs with matching tier to stakes, not habit and avoiding over-engineering and under-resourcing.
You must pick tiers for a two-page ambiguous strategic recommendation with major consequences and a fifty-page routine compliance-log reformatting job across hundreds of entries. Which routing follows the decision logic?
People also ask
How do I choose between Haiku, Sonnet, and Opus?
Should model choice depend on document length?
Can I reuse yesterday model choice for a new task?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
No videos curated for this concept yet
We are still curating the best official and community videos for this topic.
Official prep for this domain
Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.