I build a tool that measures what AI coding costs, so I’d better be able to say what mine costs. Here are my last 30 days of Claude Code: every request, priced at Anthropic’s list prices, from October 10 to November 8, 2026.

The short version: $3,435.15 across 42,861 requests. That’s an average of $114.50 a day, and nearly half of it came from subagents.

One caveat up front: I’m on a flat-rate Claude subscription, so this isn’t my bill. It’s what the same work would cost at API list prices. That’s still the number that matters if you’re deciding between a subscription and the API, or budgeting for a team that bills by the token.

How I measured it

Everything here comes from data on my own machine. Claude Code writes every request to a local transcript, with exact token counts for input, output, cache reads and cache writes. PromptConduit’s CLI prices each request against a public rate card and keeps a local log. Then:

promptconduit cost history --days 30

That’s the total. The breakdowns below come from the same log. Nothing in this post is an estimate from averages: every request is counted once, at its own token counts.

Where the money went

	Cost	Share
Main conversation	$1,786.22	52%
Subagents	$1,648.93	48%
Total	$3,435.15

Almost all of it was one model. Opus 5.5 handled 96.7% of main-conversation spend and 95% of subagent spend. Fable 5.1 covered most of the rest.

The work was spread across 14 repos: a mobile game, a few websites, and PromptConduit itself. The busiest repo took 25% of main-conversation spend.

1. Subagents spent half the budget

This surprised me most. In 30 days Claude Code launched 557 subagents, and they cost $1,649.

• The median subagent cost $1.56. The 90th percentile cost $7.21, and the most expensive single one cost $46.
• The top 10% of subagents accounted for 43% of subagent spend.
• By type, the gap is wide:

Subagent type	Count	Total	Average
general-purpose	450	$1,386.45	$3.08
fork	34	$201.38	$5.92
code-reviewer	51	$35.32	$0.69
Explore	16	$13.95	$0.87

General-purpose subagents start with their own fresh context and run on the main model, which for me is Opus. They’re the right tool for real work, but they’re easy to reach for when a cheaper agent would do. The read-only Explore agent did its job for under a dollar on average.

What I’m changing: I’m defining my common subagent tasks (search, review, summarizing logs) as custom subagents with a cheaper model set in their definition, and saving general-purpose agents for work that needs Opus.

2. Prompt caching saved me about $17,000

Here’s what a main-conversation dollar bought:

Component	Cost	Share
Cache reads	$901.25	50%
Cache writes	$634.78	36%
Output	$248.74	14%
Fresh input	$0.22	0%

Claude Code re-sends the whole conversation on every request, but almost all of it is served from the prompt cache. Over the month that came to 4.49 billion cache-read tokens. At Opus 5.5’s cache-read rate of $0.20 per million, they cost about $900. At the normal $4 per million input rate, the same tokens would have cost about $18,000.

So caching isn’t an optimization at this scale. Without it, long agentic sessions would be priced out entirely.

3. Coming back to a long session is expensive

The ten most expensive single requests all looked the same: a few hundred to a few thousand output tokens, and 480,000 to 930,000 tokens written to the cache. Each cost between $5.55 and $9.71.

That’s a cold cache. Claude Code caches context for an hour. Come back to a long session after lunch and the next request re-writes the whole context at $8 per million tokens (the 1-hour cache-write rate for Opus 5.5), instead of reading it at $0.20.

This month I had 117 of these cold-cache requests, costing $261, which was 15% of main-conversation spend.

What I’m changing: after a long break, I start a fresh session with a short summary instead of resuming a huge one. Running /compact before stepping away also helps.

4. A few sessions cost most of the money

I ran 92 sessions with main-conversation spend this month. The median one cost $10.43. The 90th percentile cost $52, and the most expensive cost $192.

The top 10% of sessions accounted for 44% of main-conversation spend. Average cost per session hides this. A handful of long, sprawling sessions decide the bill.

5. Daily spend swings a lot

• Busiest day: $341.83 (October 10), of which $212.71 was subagents
• Quietest day: $6.05 (October 26)
• Median day: $81.64. On 14 of the 30 days I spent more than $100.
• First four weeks: $1,130, $652, $412, $1,071

If you’re budgeting for a team, plan around the busy weeks, not the average. And watch for spikes as they happen, not when the invoice arrives.

What I’d tell a team lead

1. Look at subagent spend separately. It was half my total, and it doesn’t show up as “the conversation”.
2. Expect a long tail. A few sessions, a few subagents and a few cold-cache requests drive a big share of the bill.
3. Measure per repo and per person before setting budgets. Averages will mislead you, so start from real numbers.

See your own numbers

The cost tracking in this post is free and runs entirely on your machine:

curl -fsSL https://promptconduit.dev/install | bash
promptconduit init
promptconduit cost history --days 30

If you lead a team, PromptConduit shows the same breakdown for everyone: spend per member, repo, client and pull request, with budgets and alerts. It’s free for up to 3 people.