Ask the Phase, Not the Model

Open Claude, ChatGPT or Gemini for a task that actually matters, and the first choice you make usually has nothing to do with the task itself. It’s a dropdown, a model picked by name, the way you’d pick a bottle of wine you don’t know how to read: going with the one that sounds like it costs more.

Picking by name isn’t unreasonable. It just answers the wrong question. What actually decides whether a task goes well is simpler: whether you’re still figuring out how to do the thing, or you already know and you’re just doing it again.

The two moments

Every task you hand an AI system sits in one of two moments, and they call for opposite settings.

The first is exploration. You don’t have a method yet. You’re weighing conflicting instructions, working out what a good answer even looks like, before you can reliably ask for one. The second is execution. You already worked that out, maybe minutes ago, maybe last week, and now you’re running roughly the same process again on new input.

Most people default to the first setting for everything, out of a reasonable but misplaced caution: pick the strongest option, every time, just in case. Defaulting that way is expensive in the wrong direction, for a reason both major model makers document directly.

The mechanism, in the companies’ own words

Anthropic’s own platform docs describe an effort setting that controls how many tokens Claude spends before it answers, and its guidance is direct: “adjust effort based on task complexity. Simple queries may warrant low effort while agentic coding and complex reasoning benefit from high effort” (Anthropic, Claude Platform Docs, “Effort,” 2026). The same page is explicit that spending more isn’t automatically safer. Effort, it says, is “a behavioral signal, not a strict token budget,” and one of Anthropic’s own per-model tables warns that its highest setting “adds significant cost for relatively small quality gains” and “can lead to overthinking” on some tasks.

OpenAI’s reasoning-model guide draws the same line from the other model maker. Its default, medium effort, is described as “the default configuration for most workloads, and a well-balanced point on the pareto curve of latency, performance and cost” (OpenAI, “Reasoning models,” 2026). Its highest levels carry an explicit warning attached: “only use when your evals show a clear benefit that justifies the extra latency and cost.”

Read together, the two companies are saying the same thing from opposite ends. Effort, and by extension model tier, is inference-time compute you’re choosing to spend or save, not a badge of seriousness, and the right amount depends entirely on which of the two moments the task is actually in.

The clearest real version of the rule

Anthropic’s own research team ran into this exact fork building the system behind Claude’s Research feature, and the shape of their answer is the cleanest working example available.

A lead agent receives the query and, in Anthropic’s words, “analyzes it, develops a strategy, and spawns subagents to explore different aspects simultaneously,” thinking through the approach before anything else happens (Anthropic, “How we built our multi-agent research system,” 2025). The lead agent ran on Claude Opus 4, Anthropic’s most capable model at the time. The subagents it spawned, each handed one scoped, already-defined piece of the plan, ran on the faster Claude Sonnet 4, independently searching and reporting back.

The split follows the same explore-versus-execute line drawn by the effort documentation, applied inside one working system, rather than a cost-cutting afterthought: the moment of deciding what to do gets the more capable model, and the moment of doing a well-defined next piece gets the faster one. The stakes of getting that split wrong aren’t small, either. A multi-agent setup like this one runs on roughly 15 times the tokens of a single chat interaction, and Anthropic reports it outperforming a single top-tier model working alone by 90.2% on their internal research evaluation. Spend that multiplier on genuine exploration and it buys real capability. Spend it running an already-solved process expensively out of habit, and it’s 15 times the cost for the same answer you’d have gotten cheaper.

Worth separating from this: how much of a task to hand over to an agent in the first place is its own decision, made up of several smaller ones, not the question this piece is answering. In the research system above, the lead agent’s plan already fixed the target. Only the depth of each pass, explore or execute, was still a choice.

The two-question test

You don’t need Anthropic’s infrastructure to use the same split. Two questions do the work.

The first: am I still working out the approach, or running one I already trust?

Still deciding what the right approach even is? Reach for the strongest model, higher effort. The task needs judgment you haven’t formed yet.

Running a method you’ve already validated? A faster, cheaper model, lower effort earns its keep here. The judgment is done, you’re paying for execution now.

Not sure which one this is? Start high and check the second pass. If a fast model handles the repeat fine, that’s your answer for next time.

The second, for whenever the first answer isn’t obvious: would I be surprised by the shape of the output? Surprised means you’re still exploring, whatever the task looked like last time. Not surprised means you’re executing, and a faster setting is worth trying. Both companies back the same habit of checking rather than guessing. Anthropic’s own guidance says to “test your use case: the impact of effort levels varies by task type. Evaluate performance on your specific use cases before deploying” (Anthropic, Claude Platform Docs, “Effort,” 2026), and OpenAI’s own highest-tier warning, already quoted above, says the same thing from its side of the market.

A wrong guess in either direction is cheap to catch this way. A fast model on a task that turns out to still be exploration reads noticeably thinner, and a strong model on routine execution just costs more for the same answer, but only if the second question actually gets asked. And the model is only one dial on that cost: how much each session carries is another, with a routine of its own.

Neither call is permanent. A recurring task can turn genuinely novel again the day an input breaks the pattern you built the process around, and a task you were sure needed the frontier model can turn out fine on the fast one once you actually check. Run the two-question test as a habit, not a setting you fix once.

Choosing a model tier is a different question from picking which AI provider to bet on as the market reshuffles, a decision with its own framework built around switching costs and who owns the re-evaluation. That’s a strategic, once-a-year call about which vendor’s roadmap you’re committing the organization to. The question here runs on a shorter clock: a daily call inside whichever tools you’ve already chosen, made task by task, not once and forgotten.

The model names will keep changing. A new release lands and the dropdown gets one more unfamiliar entry, or an old one quietly disappears. What doesn’t move is the question underneath it: whether you’re still working out how to do this, or already running something you trust. Answer that first, and the name in the dropdown stops being a guess and starts being a formality.