Quick answer
Claude Code records a model identifier on every assistant message in its local session files, so per-model token usage is already on your machine. Split input, output, cache reads, and cache writes by model, and you can see whether Opus turns are driving your weekly cap while Sonnet does the routine work.
Where the per-model numbers live
Claude Code writes JSONL session files under your home directory, one record per message. Each assistant record carries the model name plus a usage object with input tokens, output tokens, cache read tokens, and cache creation tokens. Summing that object by model produces the split — no API key and no account required.
| Field | What it counts | Why it matters |
|---|---|---|
| model | Exact model identifier | The grouping key |
| input_tokens | Fresh prompt tokens | Grows with files in context |
| output_tokens | Generated tokens | Priciest per token |
| cache_read_input_tokens | Reused prefix | Cheap, often the largest count |
| cache_creation_input_tokens | Prefix written to cache | One-off surcharge |
Cache fields explain most of the surprise in Claude totals. A session with heavy cache reuse can show millions of tokens while costing far less than the raw number suggests.
How the split usually looks
Most working setups follow a pattern: Sonnet carries the bulk of turns, Opus handles the hard reasoning, and Haiku does short, cheap work such as summarising or classification. The interesting figure is not the total but the share each model takes of output tokens, because output is what agent runs accumulate fastest.
- Opus turns are fewer but heavier per call.
- Sonnet usually owns the largest turn count.
- Haiku rarely dominates volume but can dominate request count.
- A model you did not choose often appears through subagents or background tasks.
Read the split against your plan
- Pick the period you care about: today, 7, 30, or 90 days.
- Group tokens by model and note the share of output tokens.
- Compare the split with the tasks you actually did.
- If Opus share is high on routine work, change the default model.
- Re-check after a week to confirm the change took effect.
The split is also the honest way to explain a blocked week: it shows whether the cap went to hard problems or to a long loop on the most expensive model. UsageAtlas reads those local files and shows the per-model breakdown next to reported plan usage, so the two views line up. See the token usage report guide for the full field list, and Claude session vs weekly limits for which window a heavy model exhausts first.
Questions, answered
Does Claude report token usage per model?
Yes. Each assistant message in the local session file records a model identifier and a usage object, so per-model totals can be computed locally.
Which Claude model uses the most tokens?
Usually the one you select by default or the one subagents inherit. Opus consumes more per call, so a small turn count can still be the largest share of output tokens.
Do cache reads count toward my limit?
They are recorded separately and billed at a lower rate, but they still appear in usage totals. Session files also carry a working directory, so you can group by project and then by model inside each one before reading the cache fields.