Fifty-five billion tokens

19 Aug 26

ai

Every machine I work on ships its agent session logs to my home server: Claude Code since July 2025, and before that Zed’s agent panel, which recorded the same usage numbers.1 Each assistant message in there carries a usage block recording what it cost to produce. Fifteen months of that comes to 23,851 transcript files and 532,027 API responses.2 Fifty-five billion tokens—that’s a big number.

But mostly I just wanted to look at it.

noneunder 16M16-49M49-158Mover 158MEvery day of agent useTokens per day, 11 May 2025 to 20 August 2026. Unfilled cells are days with no session at all.JunJulAugSepOctNovDecJan 26FebMarAprMayJunJulAugMonWedFri

The gap in January is a holiday. Everything after February is a change in how I work.

Sonnet 3.7Sonnet 4Opus 4Opus 4.1Sonnet 4.5Haiku 4.5Opus 4.5Opus 4.6Sonnet 4.6Opus 4.7Opus 4.8Fable 5Sonnet 5Opus 5Jul 25Sep 25Nov 25Jan 26Mar 26May 26Jul 26The rise and fall of each modelTokens per day, 7-day rolling mean. One panel per model, ordered by first use; all panels share the same scale.

Fourteen models in fifteen months, each taking over from the last.

slop-university · 15.6Bllms-unplugged · 4.1Bdotfiles · 3.3Bextempore · 2.9Bcomp4020 (workspace) · 2.6Bcomp4020-agentic-coding-studio · 2.0Bbenswift.me · 1.9Bout-of-office-cv · 1.8Bblowing-smoke · 1.8Bcomp4020/website · 1.6Bastro-theme-anu · 1.4Bstrproxy · 1.2Bslop-salon · 1.1Bpanic-tda · 0.9Baps-ai-transparency-tracker · 0.8Bpanic · 0.8BJul 25Sep 25Nov 25Jan 26Mar 26May 26Jul 26Tokens by projectTokens per day, 7-day rolling mean, for the 16 projects I spent themost tokens on. Each panel is scaled to its own peak, with the total in the label;together they are 78% of all tokens.

Sixteen projects, three-quarters of the total.3

Jul 25Sep 25Nov 25Jan 26Mar 26May 26Jul 260204060billions of tokens0k10k20k30k40k50kindicative API cost (US$)Cumulative since day zeroThe same 467 days, counted two ways. Separate panels, separate scales.

About 86% of the total sits in the last five months.

0:003:006:009:0012:0015:0018:0021:00024681012million tokens per hourMon-FriSat-SunThe working day, as seen by the token meterAverage tokens per hour of the day (Australia/Sydney), across 334 weekdays and 133 weekend days.

The 6pm cliff is dinner and bedtime. The overnight floor never quite reaches zero, which is the scheduled agents working while I’m asleep.4

Tokens aren’t fungible, since a cached input token costs a fiftieth of what an output token does. So here is each model twice, by volume and by what it would have cost at list API prices:

Sonnet 5Opus 4.8Opus 5Opus 4.7Opus 4.6Fable 5Opus 4.5Opus 4.1Sonnet 4.5Opus 4Haiku 4.5Sonnet 4.6Sonnet 4Sonnet 3.7billions of tokensindicative API cost (US$)0246810120k2k4k6k8k10kTokens and indicative cost, by modelSame ordering in both panels, with independent scales.

Sonnet 5 has the largest token count by a distance and sits sixth on cost. Opus 4.1 and Opus 4 run the other way, at 7% of the calls and 11% of the money.

0%5%10%15%20%25%30%35%40%45%50%55%60%65%70%75%80%85%90%95%100%cache readcache write (1h)cache write (5m)outputfresh inputshare of tokensshare of spendWhere the tokens go, and where the money goesShare of the 55.9B tokens against share of the indicative US$51k.

Ninety-six per cent of those 55 billion tokens are cache reads: the same conversation prefix, read back once per turn of the agent loop. Output, the part that actually writes the code, is 0.55%.

The whole lot comes to about US$50,000 at list prices, over half of it cache reads.5 Uncached, the same work would have run to roughly $289,000.

The charts are all Vega-Lite, rendered to static SVG when I wrote this. The specs, the data behind them, and the scripts that walk the transcripts are all there.

Footnotes

  1. Zed kept per-request token counts, models and project paths for its agent panel, so those months needed no more than a format conversion to sit alongside the Claude Code transcripts. The year before that — Zed’s older assistant panel, back to April 2024 — kept the conversation text and nothing else: no usage, no model. It is in the log archive with every figure marked as an estimate, and it stays off these charts, which count only numbers something actually reported. ↩

  2. Deduplicated on the assistant message id, which matters more than it sounds: resuming a session from a different working directory files the same transcript under two project directories, and 455 files are duplicated that way. The count also includes 10,253 subagent transcripts, about 21% of the responses and 7% of the cost. I keep those out of my session counts, since one session fanning out to twelve agents is still one session. ↩

  3. The COMP4020 work still lands in three of those entries, since the studio repo and the course website are separate checkouts from the workspace directory that holds everything else. Added together they would sit second on the list, above llms-unplugged. Below the sixteen is a long tail of worktrees and one-off directories, none of it individually worth a row. ↩

  4. Which is also why that floor is so flat. A cron job has no opinion about what time it is. ↩

  5. Current list prices applied retrospectively, so it’s an indicative figure and not a bill I paid: this all ran on a Max 20x subscription at a small fraction of the price. Transcripts written before Claude Code recorded the cache TTL split are counted at the cheaper five-minute write rate, which means the true number is a little above $50k, if anything. ↩

Cite this post
@online{swift2026fiftyFiveBillionTokens,
  author = {Ben Swift},
  title = {Fifty-five billion tokens},
  url = {https://benswift.me/blog/2026/08/19/fifty-five-billion-tokens/},
  year = {2026},
  month = {08},
  note = {AT-URI: at://did:plc:tevykrhi4kibtsipzci76d76/site.standard.document/2026-08-19-fifty-five-billion-tokens},
}