The corpus, drawn.
Every figure is counted from this site's database: the changelog entries the pipeline wrote, captured system prompts, names mined from each build, diffs of Anthropic's docs and feature gate readings. Nothing is smoothed or filled in.
Figures 01 to 06 cover the newest 120 releases with a changelog, and 07 to 11 each cover their own range, which is stated under the figure. A release that's missing a value is left out of that figure.
Entries per release
120 releasesEach release gets a column, oldest on the left, as tall as the number of changelog entries the pipeline wrote for it. The dashed line is the mean, rounded down to 219. The tallest is v2.1.281 with 1,240.
Releases per week
21 weeksWeeks run Monday to Sunday, and each column counts the releases npm published that week, by publish date in UTC. A week with no release still gets its slot, at zero. 120 releases over 21 weeks, and the busiest was the week of 13 Jul 2026 with 8.
Tier mix per release
25,777 tiered, 507 notEach release's column is split by the tier of its entries and fills to 100% of the ones that have a tier. Every column is the same height, so check figure 01 for how big each release was. Across all 25,777 tiered entries, 37% are under the hood and 7% are usable the day they land, rounded down. 507 entries have no tier and are left out, and a release with none tiered at all is drawn as a plain pale column.
Did Anthropic write it down?
120 releasesThese releases share one bar, split by whether Anthropic listed them, and each segment is as long as the number of releases in it. 115 of 120 have their own entry in Anthropic's changelog. The 4 Anthropic never listed:
- Unsettled
- 1 of the 120 release is still unsettled, because the check is pending or never ran. They get a segment of their own.
What a change is
26,378 classifiedEach kind of change gets a bar, as long as that kind's count summed over the release cards of these 120 releases. They're counted off the cards, so the total won't match the entry count in figure 01. 84% are features and 8% are fixes, rounded down.
What a read cost
61 of 120 record oneA column is what the pipeline spent on model calls to write one release's changelog, summed over its stages, in US dollars to the nearest dollar. Only 61 of these 120 releases have a record, so $1,041 covers those releases and no others. A release with no record is left out. Where part of a record was lost the total is a floor, so the real cost was at least that. The three dearest: v2.1.248 $109, v2.1.281 $78, v2.1.283 $56.
How large the system prompt got
173 capturesClaude Code sends a system prompt before you type anything, and it's different for each model. Each line follows one model alias through the captured versions, 2.1.119 to 2.1.319, and its height is the prompt's size in characters. The points are evenly spaced by capture, so the step between two of them can cover one version or several. The captures come from the interactive CLI in a sandbox with an empty home directory, a wiped environment and a fake API, so the CLI starts with no settings, no CLAUDE.md and no memory. Your own setup can send a different size. The switch picks the system text or the tool schemas. Five aliases are drawn at first, and the chips add the rest.
- default 25,973 → 5,877
- opus 26,083 → 6,144
- sonnet 26,069 → 5,877
- haiku 26,078 → 8,720
- opusplan 26,069 → 5,877
- claude-3-5-haiku 26,028 → 27,452
- claude-3-5-sonnet 26,030 → 27,452
- claude-3-7-sonnet 26,030 → 27,452
- claude-3-opus 25,979 → 27,401
- claude-opus-4-0 26,083 → 6,144
- claude-opus-4-1 26,083 → 6,144
- claude-opus-4-5 26,062 → 27,452
- claude-opus-4-6 26,062 → 27,452
- claude-sonnet-4-0 26,068 → 27,452
- claude-sonnet-4-5 26,070 → 27,452
- claude-sonnet-4-6 26,069 → 27,452
- sonnet[1m] 25,989 → 27,239
- characters of system text, 2.1.119 to 2.1.319
- default 50,296 → 19,064
- opus 50,322 → 18,754
- sonnet 50,300 → 19,064
- haiku 50,298 → 19,064
- opusplan 50,300 → 19,064
- claude-3-5-haiku 50,298 → 67,934
- claude-3-5-sonnet 50,300 → 67,934
- claude-3-7-sonnet 50,300 → 67,934
- claude-3-opus 50,296 → 67,681
- claude-opus-4-0 50,322 → 18,754
- claude-opus-4-1 50,322 → 18,754
- claude-opus-4-5 50,296 → 33,628
- claude-opus-4-6 50,296 → 33,628
- claude-sonnet-4-0 50,296 → 33,628
- claude-sonnet-4-5 50,300 → 33,628
- claude-sonnet-4-6 50,300 → 33,628
- sonnet[1m] 50,296 → 33,628
- characters of tool schemas, 2.1.119 to 2.1.319
- Left out
- (unset), claude-fable-5, claude-haiku-4-5, claude-mythos-5, claude-opus-4-7, claude-opus-4-8, claude-opus-5, claude-opus-5[1m], claude-sonnet-5, claude-sonnet-5[1m]. Each is missing a capture for at least one version in this range, and a line would drop to zero there as if the prompt had vanished.
- Scope
- The Claude Code CLI only. The Agent SDK's prompt is captured too, and it's on the prompts ledger, since it belongs to a different program.
How much there is to know
159 releases minedEvery environment variable, feature flag, setting, command line option and tool name the miner found in each build. It's counted from the code, so a name shows up here whether or not Anthropic ever documented it.
The names present in each build, one column per release, stacked by family. The legend gives the newest release's counts.
What each release added sits above the rule and what it removed sits below, both on the same scale. 3,646 added and 1,898 removed in all.
- Range
- 159 releases, 2.1.138 to 2.1.319. Only releases the miner read for all five families are drawn. A release read for one family would look as if the other four had vanished.
Where each release spent its weight
750 areasEach release's changelog entries, grouped by the product area they touch. Each is drawn as a share of its release, because releases differ a lot in size and the question here is what each one was about.
A release's column fills to 100% of its entries that have an area. The eight largest areas across the range get a band each and the rest share "Everything else". Legend numbers are entries per area across the range.
The same entries by tier, counting the ones that have a tier.
- Range
- 25,777 entries across 88 releases, 2.1.219 to 2.1.319, in 750 distinct areas. Only entries with an area label are counted. The current pipeline writes that label and older releases don't have it, so they're left out of this figure. "Everything else" only holds labelled entries outside the top eight.
How often Anthropic rewrites their own docs
281 daysEvery page of Anthropic's published docs, polled and diffed. Each square is a day, shaded by how many page changes were recorded that day, in the five bands keyed below. A page changed twice in a day counts twice. A change is dated by the docs' own last edit time where they publish one, otherwise by when this site saw it, as a UTC day.
23,237 page changes over 281 days. 14,245 of them were rated significant. The busiest day had 1,099.
The same days again, with the lines those changes added above the rule and the lines they deleted below it, on one scale. One big rewrite and a pile of small edits can shade a day the same, and this is where they come apart. 2,276,198 lines added and 1,560,802 deleted in all.
The same page changes by docs source.
- Range
- 2026-01-30 to 2026-11-06. That's every change the differ has recorded since it started. A day with no change is drawn as an empty square.
- Left out
- The differ's starting reads, where it had nothing to compare against: 1,109 changes and 1,185,902 added lines. A read like that records every page as new, so all it marks is the day this site started reading. Only those reads are left out, and real edits on the same day still count.
What was switched on when
352 readingsClaude Code reads its feature gates from a service at startup, so a build's behaviour depends on more than its code. Every gate read during a release's own pipeline run lands in that release's column, stacked by state, so a column is as tall as the number of gates read. Legend numbers are readings summed over all releases.
- Account delta
- At most 1 gate in any one release read differently for the pipeline's account than for an anonymous one. The service answers per account, so every reading drawn here is that account's.
- Range
- 19 releases, the ones whose own run took a gate reading. There's also an older snapshot, taken on one evening and attached to many earlier releases at once. It's left off because it would pass one evening's reading off as their history. What a gate reading is and is not.