How to Get Better Answers From Claude
at a Lower Cost

Claude can save your team hours every week, but how you use it decides both the quality of the answer and what it costs. These are the habits that make the biggest difference day to day.

What Drives Token Consumption

1. Pick the smallest model that can do the job

Your model choice moves cost more than anything else. Running everything on the largest model available is the single biggest consumer of tokens and the largest avoidable cost.

• Haiku. Extraction, classification, routing, short summaries. Use it when the answer already sits in what you sent and the work is finding it.
• Sonnet. Your default for real work. Drafting, analysis, reasoning across a set of records.
• Opus. Deep synthesis across a lot of material, and judgment calls where being wrong is costly.
• Fable. Frontier tier, reserved for the hardest problems. Rarely the right choice for anything but extensive development work.

Anthropic releases new versions of each model regularly, so use the latest version available in whichever tier fits the task.

2. Turn extended thinking off unless the reasoning is the deliverable

Thinking mode adds reasoning tokens to every call. Leave it off for lookups, summaries, and short drafts.

3. Start a new chat for every task

Each turn resends your entire conversation. The same question asked at message thirty costs several times what it cost at message three.

4. Name the record you need

Say you want to know where a renewal stands and the data is in Salesforce. You could ask what’s at risk across your book of business, or you could ask what happened on the Acme renewal in the last thirty days. The first request loads hundreds of records, while the second loads one. Point at the account, record, or report you actually mean.

5. Send the excerpt, not the document

If you’re asking about an attachment, sending a full document such as a contract to ask about four clauses means paying to read every page you didn’t need. Paste the section instead.

6. Be clear about what you want back in your first request

Tell Claude how you want the output: the length, the format, and the audience. Three bullets, one paragraph, a table with the columns you name. Output costs more per token than input, so an open-ended request gets you a long answer you then have to trim.

7. Give all the context in your first message

One complete prompt costs less than eight rounds of correction, because every correction reprocesses everything before it.

8. Ask for an edit, not a rerun

Request the change to the paragraph you want fixed instead of regenerating the whole output.

9. Don’t trust it with math

Language models predict text; they don’t calculate. Totals, percentages, and date arithmetic come back wrong often enough that you can’t rely on them. Get numbers from the source system and use Claude for the words around them.

Other Tips

Always include context, task, inputs, format, and constraints. Who you are and what you’re working on, what you want done, the material to work from, the shape of the answer, and the limits. People often skip two things: saying what to leave out, and telling Claude to ask questions rather than guess when something is missing.

When you do the same task twice, make it a skill. If you find yourself writing the same kind of prompt every week, stop rewriting it. A skill holds the instructions, the format, and the examples so you don’t rebuild them from scratch each time. You get more consistent output, you stop paying to re-explain the same context, and the next person who needs that task gets your version instead of inventing their own. Rolling Claude out across your team? Our Claude consultants can help you set up the models, skills, and habits that keep answers sharp and costs predictable. Talk to a CloudMasonry consultant.