Claude Code in October: 8.4 Billion Tokens, 93% Cache, $4,323 at List Price
Updated November 3, 2026 · first published November 3, 2026
Between 1 October and 3 November 2026, Claude Code on one machine recorded 70,592 unique assistant messages and 8.4 billion tokens, of which 93% of input-side tokens were cache reads. At Anthropic's API list prices that was about $4,323: roughly $3,636 if every cache write were 5-minute, $4,735 if every one were 1-hour. Cache reads were most tokens and 28% of the value.
Published 3 November 2026 · Claude Code transcripts, 1 October to 3 November 2026 (UTC; 1 to 3 November partial) · list prices, not a bill
This is the Claude Code counterpart to our Codex Pro $100 month. The two tools use different providers, plans and meters, so compare the method, not the totals. The counting method, including why raw log sums are 2.5x too high, is in Claude Code logs overcount tokens.
Tokens versus dollars
Reading cached context is cheap per token, so it is a large share of tokens and a modest share of dollars. Output and cache writes are the opposite: a small token share, a large value share. Cache writes alone were 58% of the value on 561 million tokens, so what you write into the cache matters as much as what you read from it. 42% of the cache-write tokens were 1-hour writes, which list at 60% more than 5-minute writes for the models we used most (Opus 5.5 $8 against $5, Sonnet 5.5 $4 against $2.50). For the mechanics, see prompt cache attribution.
By model
| Model | Messages | Output tokens | List value |
|---|---|---|---|
| Claude Opus 5.5 | 38,379 | 21.8M | $3,130.25 |
| Claude Sonnet 5.5 | 25,440 | 12.1M | $1,039.98 |
| Claude Fable 5.1 | 223 | 0.6M | $96.30 |
| Claude Haiku 5.5 | 6,300 | 3.8M | $37.36 |
| Claude Opus 4.8 | 83 | 0.2M | $16.39 |
| Claude Haiku 4.5 | 167 | 0.1M | $2.50 |
About 29,700 of the unique messages came from sub-agents (sidechains), which share the same log files; they are included.
Day by day
What would routing change?
Hypothetical arithmetic, not a recommendation. On Anthropic's list prices, Sonnet 5.5 costs exactly half of Opus 5.5 in every category (input $2 against $4, cache read $0.10 against $0.20, output $10 against $20). If a share of the Opus 5.5 work could run on Sonnet 5.5 with the same acceptance rate, the list value falls like this:
| Share of Opus 5.5 work moved | Saving at list price | New total |
|---|---|---|
| 10% | $157 | $4,166 |
| 25% | $391 | $3,931 |
| 50% | $783 | $3,540 |
The assumption does all the work. The acceptance rate and the retry rate are unknown until you test on your own tasks, and a cheaper model that needs more retries can cost more. Measure cost per successful task before you move anything.
How we measured it
- Claude Code transcripts on one Mac, assistant entries with token usage, 1 October to 3 November 2026 UTC, deduplicated by message id plus request id keeping the maximum of each field (see the method article).
- Prices from Anthropic's pricing page, checked 3 November 2026. Opus 5.5: $4 input, $5 or $8 cache write (5-minute or 1-hour), $0.20 cache read, $20 output per million tokens. Sonnet 5.5: $2, $2.50 or $4, $0.10, $10. Haiku 5.5 up to 100,000-token prompts: $0.10, $0.125 or $0.20, $0.01, $0.50, with higher rates over 100,000. Fable 5.1 and the older Opus 4.8 and Haiku 4.5 are priced from the same page.
- Cache writes use the 5-minute and 1-hour token counts recorded in the logs (324 million and 237 million tokens). The range in the first paragraph prices all writes at one rate.
What the numbers leave out
- List-price value is not a bill. If you are on a flat Claude plan, this is an API-equivalent figure only, and we do not know of a way to read the real charge from the logs.
- One machine only. The window ends mid-day on 3 November, and our own live session kept adding lines while we measured.
- Haiku 5.5's long-prompt rate is applied per message from its input-side tokens, which is an estimate.
Frequently asked questions
How many tokens does Claude Code use in a month?
On our machine between 1 October and 3 November 2026 it was 8.4 billion tokens across 70,592 unique messages, 93% of input-side tokens being cache reads. Yours depends entirely on your workload.
How much would that cost at API prices?
About $4,323 at Anthropic's list prices, between $3,636 and $4,735 depending on whether cache writes are priced at the 5-minute or 1-hour rate. It is list-price value, not what a subscription charges.
Why are cache reads most of the tokens but not most of the cost?
Cache reads are priced at a fraction of normal input (for Opus 5.5, $0.20 per million tokens against $4 base input). They were 93% of tokens and 28% of value, while cache writes were 58% of value on 6.7% of tokens.
Would moving Opus work to Sonnet halve the bill?
Only for the work you move, and only if the results are as good. Sonnet 5.5 lists at half the Opus 5.5 price in every category, but retries and rejected work can erase that. Test cost per successful task first.
Related
- Claude Code logs overcount tokens 2.5x
- How to check and reduce Claude Code token usage
- Claude vs Codex in October 2026
- We used Codex Pro $100 for 30 days: 14 resets
Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →