What Does the Effort Setting on an AI Model Change for Marketers?

What does the effort setting on an AI model actually change for marketers?

The effort setting controls how much work a model puts into each response. Low effort is faster and cheaper but thinks less. High effort is slower and costs more but reasons more fully. For marketers, it means one model can handle quick tagging and careful analysis, as long as someone picks the right level per job.

Until recently, the main cost and quality lever in AI tools was the model itself. You picked a big model for hard work and a small one for simple work. Now both Anthropic and OpenAI expose a second lever inside the same model, and most marketing teams have not touched it.

That matters because the default setting is rarely the right one for every task. A content brief, a lead classification, and a competitive analysis do not need the same amount of thinking. Here is what the setting does, where it helps, and where it quietly hurts.

What is the effort setting in plain terms?

It is a dial that tells the model how thorough to be. Anthropic's documentation describes effort as a way to control how many tokens Claude spends when responding, trading thoroughness against token efficiency. OpenAI's API has a similar control called reasoning effort, which its guide says tells the model how much to think.

Tokens are the units models read and write, and you pay for them. More thinking means more tokens, which means more cost and more time before an answer arrives. Less thinking means a faster, cheaper answer that may skip steps a harder task needed.

Anthropic lists five levels: low, medium, high, xhigh, and max. OpenAI's guide says its supported values depend on the model and can include none, minimal, low, medium, high, xhigh, and max. The names line up roughly across vendors, but the behavior behind each name is not identical, so treat them as relative settings, not shared standards.

How is effort different from choosing a bigger model?

A bigger model has more capability built in. Effort changes how much of that capability a model uses on a given request. Choosing a model sets the ceiling. Choosing effort sets how close to that ceiling each answer goes. You can now run a strong model cheaply on easy work and push a small model harder on difficult work.

This changes how I think about model selection for marketing teams. The old question was which model fits this job. The better question is which model and effort pair fits this job at a cost the team can sustain. Sometimes a small model at high effort beats a large model at low effort, and sometimes it does not.

I covered the model side of that decision in which AI model for which marketing job. Effort sits one layer below that choice, and it is where a lot of cost and quality drift now happens.

What does effort actually change in the output?

It changes more than the visible answer. Anthropic's documentation says effort affects all tokens in a response, including the text, tool calls, and the model's thinking when thinking is active. At lower effort, Claude tends to make fewer and terser tool calls. At higher effort, it may explain its plan and make more calls.

For marketers using AI inside automations, the tool call part is the one to watch. An agent that researches a company before writing a first line may look up fewer sources at low effort. That is fine for a quick tag. It is a problem if the whole point of the step was the research.

Anthropic is direct about this for its newest small model. Its guidance for Claude Haiku 5.5 says that in long agent prompts, the model is more likely to skip a search, stop early, or skip a check at low effort, and it recommends high effort for knowledge work and strict instruction following. That is a useful warning for anyone running research steps on autopilot.

Is effort a hard limit on cost?

No. Anthropic's documentation calls effort a behavioral signal, not a strict token budget. A model at low effort still thinks on hard problems, just less than it would at higher effort. If you need a firm ceiling on spend, you need separate controls, such as a maximum output length and budget alerts at the account level.

This is a common misunderstanding. Teams lower effort expecting a predictable bill and are surprised when a few hard inputs still produce long, expensive responses. Effort shifts the average. It does not cap the worst case.

For budgeting, I treat effort as a way to tune the typical request and a hard output limit as the safety rail. Both belong in the configuration of any automation that runs at volume, next to a note explaining why each value was chosen.

Which marketing tasks suit low effort?

Low effort suits short, well-defined tasks with a clear right answer: tagging a form submission by segment, extracting a job title from a profile, labeling the sentiment of a reply, or reformatting call notes into fixed fields. These jobs repeat often, are easy to check, and rarely benefit from extra reasoning.

OpenAI's guide frames lower effort as favoring speed and lower token usage. Anthropic's best-practice notes say low effort can significantly reduce response times and costs on straightforward tasks. Both point the same direction: simple, high-volume work is where the savings are.

The test is whether a person could do the task in a few seconds without looking anything up. If yes, start at low effort and check a sample of outputs. If the sample holds up, leave it there. If you see mistakes, raise effort one level before you switch models.

Which tasks deserve high effort?

High effort suits work where reasoning quality is the product: positioning analysis, competitive comparisons, research that pulls from several sources, content briefs that must follow a detailed style guide, and any agent step that decides what to do next. These are the tasks where a skipped check costs more than the extra tokens.

Instruction following is the quiet reason to raise effort. Marketing prompts are often long, with brand rules, banned words, formatting requirements, and audience notes. At lower effort, a model is more likely to satisfy the main request and drop a rule near the bottom of the prompt. Anthropic's Haiku 5.5 guidance specifically names strict instruction following as a reason to use high effort.

I would also use higher effort for anything that publishes. A blog post, a landing page section, or an email that reaches buyers deserves the careful setting, plus human review. The cost difference on a single piece of content is small. The cost of a wrong claim in public is not.

Does changing effort affect prompt caching?

It can. Anthropic's documentation warns that changing the top-level effort value between requests invalidates prompt caching, because effort shapes the prompt the model sees. For workflows that rely on caching a long system prompt, pick an effort level per workflow and keep it constant rather than changing it request by request.

Prompt caching matters for marketing automations that reuse the same long instructions thousands of times, such as a brand voice guide or a classification rubric. Caching those instructions is one of the easiest ways to cut cost. Flipping effort back and forth can quietly undo that saving.

Anthropic's docs describe a per-message way to change effort on some newer models that preserves the cache, currently in beta. If your workflow needs different effort levels inside one conversation, check the current docs for your model before building around it.

How should a marketing team set effort levels?

Set effort per workflow, not per company. List your AI steps, label each by risk and complexity, assign a starting level, and test on real examples before rollout. Write the chosen level and the reason into your automation docs, so the next person does not reset it to the default without knowing why.

Defaults deserve a second look. Anthropic's documentation says most Claude models default to high effort, while Claude Opus 5.5 and Claude Haiku 5.5 default to medium. OpenAI's guide calls medium the default configuration for most workloads. A team that never sets effort explicitly is trusting whatever the vendor chose for the model it happens to use.

Model upgrades are the other trigger. When a vendor ships a new model, effort levels can behave differently than on the old one, so rerun your test set rather than assuming the old setting still fits. I wrote about protecting prompts through those changes in writing prompts that survive a model upgrade, and effort belongs in the same checklist.

What should you do next?

Open the configuration of every AI step your team runs and write down the model and effort level each one uses. Move simple, high-volume steps to low effort after a sample check. Raise effort on research, analysis, and anything that publishes. Keep levels constant inside cached workflows, and retest after every model upgrade.

If you are weighing a small model for your go-to-market automations, my breakdown of which GTM jobs should move to Claude Haiku 5.5 pairs well with this one.

I build websites and the automations behind them for lead generation, and settings like effort are where a lot of hidden cost lives. If you want a second pair of eyes on how your AI steps are configured, reach out at pravinkumar.co and let's chat.