Pick the lightest setting that does the job
Most everyday work doesn't need the biggest model or the most thinking. Starting light saves energy and stretches your usage limits too. Here's what to pick for common tasks in Claude and ChatGPT.
Thinking is the biggest factor
When a model "thinks" before it answers, it does a lot more work. Research has found that this can use around 30 times more energy than answering right away.
Faster, lighter models like Haiku in Claude or Instant in ChatGPT use less computing power for each answer.
Every redo or long back and forth is another round of work. A clear first ask usually saves more than any setting.
What to pick for each task
In Claude, the dropdown next to the message box sets the model and the Effort. In ChatGPT, the Chat tab has a thinking slider (Instant, Medium, High, and so on), and the Work tab also lets you pick a model.
| TASK | CLAUDE | CHATGPT |
|---|---|---|
| LIGHTQuick email, reply, or rewriteProofreading, changing the tone, a short answer | Haiku 5.5, Effort Low | Chat tab, Instant |
| LIGHTSummarize something you paste inMeeting notes, an email thread, a short document | Haiku 5.5 or Sonnet 5.5, Effort Low | Chat tab, Instant |
| MEDIUMEveryday writingA memo, letter, agenda, or announcement | Sonnet 5.5, Effort Medium | Chat tab, Medium |
| MEDIUMPull details out of a fileDates and deadlines, a list of contacts, key terms | Sonnet 5.5, Effort Medium | Chat tab, Medium |
| MEDIUMMake a fileA Word document, slides, or a spreadsheet | Opus 5.5, Effort Medium (the default) | Work tab, Default model, Medium |
| HEAVYCompare or check numbersTwo bids or proposals side by side, a budget, a spreadsheet review | Opus 5.5, Effort Medium, or High if it's high stakes | Work tab, GPT-6.1 Sol, Medium or High |
| HEAVYA tricky decision or careful reviewLots of moving parts, contract language, real stakes | Opus 5.5, Effort High | Work tab, GPT-6.1 Sol, High |
| HEAVYResearch across many sourcesFunding programs, regulations, market questions | Research, only when a normal chat can't answer it | Deep research, only when a normal chat can't answer it |
Save for the hardest jobs: Fable 5.2 and Effort Extra or Max in Claude, and GPT-6 Astra, Extra High, or Pro in ChatGPT. These are the heaviest options and also use up your limits fastest.
Small habits that add up
- Start light and move up. If the answer isn't good enough, then turn the setting up. Don't start at the top.
- Start a new chat when you switch topics. In a long chat, the model rereads everything above each time you send a message.
- Ask clearly the first time. Say who it's for, what you need, and how long. Fewer redos means less work for the model.
- Ask for shorter answers when you don't need the long version.
- Turn on web search, Research, or extra tools only when the task needs them.
How sure is this?
Neither Anthropic (Claude) nor OpenAI (ChatGPT) publishes energy numbers for its models, so there's no exact chart. This guide follows what outside research consistently shows: answers with thinking turned on use far more energy than quick answers, and smaller models use less than larger ones. Treat it as a rule of thumb, not a measurement. Sources: Hugging Face study on reasoning models and estimates of AI energy use per answer.
Model names are as of November 2026 and will change over time. The idea stays the same: pick the lighter option first.
