Which model uses most tokens is usually settled by the default, not by the price list. Capability sets the rate per token. Five factors set the volume, ranked below by how much each one moves an ordinary week.

1. Default model you start on

Every session begins with a choice, and most turns never revisit it. Subagents can inherit that choice too, so a model you did not pick can appear in the log and take a share of the total. This is the largest factor and the easiest one to change.

2. How long each turn runs

Output grows with every file read, test run and retry inside a turn. A refactor that runs unattended for twenty minutes can log more output than a day of short questions. Long turns also resend context, which leads to factor four.

3. Thinking, and the effort setting

Current Claude models think before they answer, and effort controls how much. The defaults differ: medium on Opus 5.5 and high on Sonnet 5.5, according to Anthropic's model guide. The same prompt at a higher effort produces more output tokens, and Anthropic says tuning effort is often a better lever than switching models.

4. Context and prompt cache

Context is sent again with every call, so a large one multiplies input. Caching makes the repeated part cheap: a cache hit costs 5 percent of base input on Opus 5.5 and Sonnet 5.5. Those reads still appear in token totals, so a session with heavy reuse can look enormous and cost little.

5. Model generation and tokenizer

Anthropic says Claude 4.7 and later use a newer tokenizer that produces about 30 percent more tokens for the same text than the older one. Two models can read the same repository and log different counts. That is a measurement difference rather than a spending difference, so compare cost across generations instead of raw tokens.

The ranking in one table

Rank Factor What it changes Where to look
1 The default model Every turn that never switches The model setting in the tool
2 Turn length Output, plus repeated context Session duration and files read
3 Thinking and effort Output tokens per answer The effort setting
4 Context and cache Input volume at a lower price Cache read share
5 Tokenizer generation The count for identical text The model version

What to change first

Work down the list rather than across it. Moving routine work to Sonnet halves the rate per token. Shorter turns cut output and repeated context together. Effort tuning can reduce output without a model change. The tokenizer comes last because you cannot act on it.

UsageAtlas shows tokens, models, requests and estimated cost from local logs, free and open source, so you can see which factor is moving before changing a policy.

The order shifts with the work. Two people running long unattended agents can find turn length ahead of the default, because each run keeps going and picks up whatever model its subagents use. A team that mostly asks short questions will see the default dominate every week.

One caution. Price per token alone is a poor guide. A cheap model that needs three attempts can finish behind the expensive one that needed a single pass, and the split will show that as more tokens on the cheaper tier, not less.