The result: Start every model at its own default effort level, drop to low for simple, bounded work, and step up to high, xhigh or max only when your own checks show a better answer. That rule stops you paying for thinking a task never needed. Defaults differ: Claude Opus 5.5 starts at medium, Claude Opus 5 at high.
Effort mode sets how much a model reasons before it answers. Same model, same knowledge, different thinking budget. Anthropic calls effort "a behavioral signal, not a strict token budget," so treat each level as a direction, not a fixed price.
Start at your model's default, then move one level per task
The fastest way to pick an effort level is to accept the vendor's default and move one step at a time. Here is the workflow for a new task type.

- Look up the default for your exact model. Use the table in the next section. Expected result: you know your starting level. Check: the model name matches, version included.
- Run five real examples at the default. Use tasks you actually do, not toy prompts. Expected result: a baseline you can judge. Check: would you ship these answers as they are?
- Drop one level and rerun the same five. Expected result: faster, cheaper answers. Check: if quality holds, keep the lower level.
- Step up one level only where answers failed. Expected result: fixes on the hard cases. Check: the gain is visible, not just longer thinking.
- Lock the level per task type. Write it down next to the prompt or workflow. Check: the next teammate uses the same setting.
Anthropic's own advice for Claude Opus 5.5 is to "run an effort sweep on your own evals rather than carrying settings over from an earlier model." Steps 2 to 4 are that sweep, in small.
Defaults differ by model, so check yours first
Most Claude models default to high, but Claude Opus 5.5 and Claude Haiku 5.5 default to medium, per Anthropic's effort documentation. Moved from Opus 5 to Opus 5.5 without setting effort? Your requests now run one level lower. For Haiku 5.5, Anthropic advises starting at medium.
| Model | Default effort | Levels offered |
|---|---|---|
| Claude Opus 5.5 | medium | low, medium, high, xhigh, max |
| Claude Haiku 5.5 | medium | see Anthropic docs |
| Claude Opus 5 and earlier Opus | high | see Anthropic docs |
| Claude Sonnet 5.5 | high (Claude API) | low, medium, high, xhigh, max |
| GPT-5.5 and GPT-5.6 | medium | model-dependent, from none to max |
| Gemini 3.1 Pro (preview) | high | low, medium, high |
| Gemini 3.6 Flash | medium | minimal to high |

The OpenAI reasoning guide calls medium "the best starting point" for GPT-5.5 and GPT-5.6. Supported values can run from none to max, depending on the model. GPT-6 Astra rejects none with an error. Don't copy settings across models blindly. The defaults above are for the APIs; check the settings in ChatGPT and Claude separately.
Use low or medium for simple, bounded work
Low and medium are the right levels for short, well-defined tasks with one clear answer. Anthropic describes low as "most efficient," with "significant token savings with some capability reduction."
- Quick factual questions and formatting: low.
- Summaries, rewrites and everyday writing: low or medium.
- Routine extraction into a fixed format: medium, then try low.
On Claude Opus 5, effort will not reliably trim a wordy answer. Anthropic says changing effort "does not reliably shorten responses," so ask for the length you want in the prompt.
Use high, xhigh or max only for hard multi-step work
High and above pay off when an early wrong step ruins the rest. Think debugging across files, multi-step plans and agent runs. For Claude Opus 5, Anthropic's advice is to start with high, then step up to xhigh for demanding coding and agentic work, or to max.
If you read older advice to start coding at xhigh, that was for Opus 4.7 and 4.8. For Opus 5.5, run your own sweep from medium. For max, Anthropic's Opus 4.7 guidance warns it "adds significant cost for relatively small quality gains" on most workloads and can lead to overthinking. OpenAI says to use xhigh only "when your evals show a clear benefit."
Every row starts from your model's default; the right column is where tasks of that type usually end up once your checks settle it.
| Task type | Where it usually ends up |
|---|---|
| Quick factual question, routine formatting | One level below the default: low (or none where offered) |
| Everyday writing, general help | The default |
| Debugging, multi-step planning, complex analysis | The default, or high if your checks show a gain |
| Long-running coding or agent work | High, then xhigh if your checks show a gain |
| Frontier problems | Max, only after xhigh falls short |
Why one task list needs different levels
Take two requests: a short email summary and a hunt for an intermittent crash spread across five files. One fixed level serves one of them badly.

At high, the summary pays for depth it never uses. At low, the bug hunt gets a shallow first guess. That problem needs logic traced across files. Set the summary to low and the debugging to high, and each gets the thinking it needs. This is the same idea as a model router, applied inside one model.
What breaks when effort is set wrong
Most effort mistakes come from one setting used everywhere or a default nobody checked.
- High on everything as a safety net. Fix: run step 3 and drop a level wherever quality holds.
- Low on everything because it feels faster. Fix: step up for math, planning or anything with dependent steps.
- Expecting effort to fix a vague prompt. Fix: add the missing context first. More thinking can't recover facts you never gave.
- Assuming the old default. Fix: recheck after every model upgrade. Opus 5.5 moved to medium.
- Flipping effort mid-conversation. Anthropic says changing the effort value between requests invalidates prompt caching. Fix: set effort per workflow, not per message, or you may lose cache savings. See what drives an AI agent's cost.
Make it a habit: a monthly effort check
Effort works best as a standing setting per task type, reviewed on a schedule.
- Once a month: list your top five recurring tasks and the level each one uses.
- After any model upgrade: rerun step 2 at the new default before changing anything.
- When a bill jumps: look first at tasks running at xhigh or max.
- Decision rule: default first, one level down if quality holds, one level up only when it fails.




