There’s a lot of arguing about what AI coding assistants cost and whether they’re worth it. Almost nobody posts receipts. I have receipts — my assistant and I built the reporting pipeline together, on itself.

The setup: I run Claude Code on a Max plan. Claude Code keeps complete local transcripts, including per-message token usage. So we wrote a script that sums every message across every session, prices it at each model’s actual API rates (including prompt-cache write and read rates), and publishes a daily ledger. Everything below comes from that.

One heavy day

July 7 was the heaviest day I’ve recorded: $423.21 of API-equivalent usage. Actually paid: $0 — it rode the subscription.

What that bought, in one continuous session with one agent: a live electricity cost calculator with an EIA rate pipeline that refreshes itself monthly by cron; an auth-and-storage service written in Node, then ported to Swift/Hummingbird 2 and swapped into production live (memory dropped from 54 MB to 24.7 MB — the measured comparison is published); a benchmark report site; status boards; and the beginnings of a documented platform that now creates new deployed apps in one command. All running on one Orange Pi behind a Cloudflare tunnel.

Where the money actually goes

The surprise isn’t the total — it’s the shape. The models wrote about 1.45M tokens that day. They re-read about 475M.

Every API call in an agentic session replays the conversation so far. Late in a long session that’s 300K+ tokens of context per call, thousands of calls per day. Prompt caching bills those re-reads at 10% of the fresh-input rate, and they still ended up as ~70% of the day’s cost. Output — the code, the answers, the thing you think you’re paying for — was about 20%.

If you’re evaluating AI dev tools on price, this is the number to understand: you pay for context, not code. Long sessions are quadratic-ish in cost; splitting work into fresh sessions is cheaper than it feels.

The model mix matters more than anything

Across the trailing nine days, the ledger splits like this: the big models (Opus, and Anthropic’s new Fable tier) did the architecture, debugging, and prose — a few hundred dollars API-equivalent. Haiku, the small model, handled 1,103 calls for $15.70 total. The days I leaned on it cost $2–4; days that ran everything through big models cost $100+.

Since then we’ve made it policy: mechanical work — scripted file edits, doc sweeps, bulk generation — goes to Haiku sub-agents with explicit instructions; judgment stays on the big model; anything fully deterministic becomes a plain script, model tier zero. The expensive model writes the spec, the cheap model executes it, the script does it for free next time.

The mistakes are the credibility

First version of the day report said $830. It was wrong twice: I counted whole transcript files (which swept in messages from prior days sitting in still-active sessions), and under-deduplicated streamed records. Filtering by per-message timestamps got the real number, $423. I’m leaving the correction in the published report — if you’re going to post receipts, post the audit trail too.

Is it worth it?

At subscription prices, obviously — the plan absorbed roughly its own monthly cost in API-equivalent value in one day. At raw API prices, $80–160 for that day’s output (depending on model tier) against what shipped — I’ve billed clients more for less, and so have you.

The ledger updates nightly and is public. The whole platform it measures — the playbook, the one-command deploy tooling, the scripts that generate these very reports — is open on GitHub. Argue with the numbers, not the vibes.