Entries per release
120 releasesA handful of releases carry most of the corpus: 107 entries on the average one, 876 in v2.1.248, the biggest here.
Releases per week
20 weeksA version lands most weekdays: 120 across 20 weeks, peaking at 8 in the week of 4 May 2026.
Tier mix per release
11,852 tiered, 1,082 notOf the entries that carry a tier, 38% are under the hood and only 9% are something to use the day they land. The other 1,082 entries were never tiered and are left out of this chart.
Did Anthropic write it down?
120 releases115 of 120 carry Anthropic’s own notes. The 5 releases that never got one: v2.1.124, v2.1.177, v2.1.182, v2.1.213, v2.1.242.
What a change is
13,094 classifiedNew behaviour dominates: 77% of classified changes are features, against 10% fixes.
What a read cost
24 of 120 record oneCost has only been recorded recently, so this is 24 of the 120 releases drawn here and $438, not the corpus. The spread is the point: v2.1.248 $109, v2.1.271 $31, v2.1.269 $30.
How large the system prompt got
136 captures136 releases captured, 2.1.119 to 2.1.278, 17 model aliases each.
- default 25,973 → 6,051 characters of system text
- opus 26,083 → 9,431 characters of system text
- sonnet 26,069 → 27,377 characters of system text
- haiku 26,078 → 27,479 characters of system text
- opusplan 26,069 → 27,377 characters of system text
Claude Code sends a stock system prompt before you type anything, and a different one per model. This is its size in characters, captured from the request the client actually sent, 2.1.119 to 2.1.278. Five aliases are drawn to begin with; the rest are a click away.
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]. A capture is missing for at least one release in this range, and a line chart draws a gap as a zero, which would read as a prompt that briefly vanished.
Scope Claude Code itself. The Agent SDK's prompt is captured beside this one and is a different program; drawing both on one axis would invite a comparison neither supports.
How much there is to know
122 releases minedEvery environment variable, feature flag, setting, command-line option and tool name the miner found in a build. This is not documentation coverage; it is what exists.
Range 122 releases, 2.1.138 to 2.1.278. Only the releases the miner read for all five families are here. A release it read one family for would otherwise draw as the moment the other four disappeared.
Where each release spent its weight
553 areasEach release's entries, grouped by the part of the product they touch, as a share of that release. A share and not a count: the busiest build in this range carries several times the entries of the quietest, and the question here is what a release was about rather than how large it was.
Range 11,852 entries across 51 releases, 2.1.219 to 2.1.278, in 553 distinct areas. Only releases the current pipeline labelled are here: an area is written by the deep run, and everything read before that run existed carries none. An unlabelled release is not an "everything else", it is a release nobody asked the question of.
How often Anthropic rewrites their own docs
40 days6,886 page changes over 40 days, 2026-08-14 to 2026-09-22. 4,361 of them were read as significant.
Busiest day: 670 pages.
1,225,859 lines added, 952,881 deleted, across 40 days.
Every published page, polled and diffed. The squares count pages changed in a day; the lines underneath count the lines behind those changes, because a day of one enormous rewrite and a day of thirty one-line edits look nothing alike.
Range 2026-08-14 to 2026-09-22. The differ started then, so this is the whole record and not a window into a longer one. Days with no change are drawn as empty rather than skipped.
Left out The differ's first run on each source, 1,109 changes and 1,185,902 added lines. Every one of them is an addition and not one is a deletion, because that run was the differ reading the corpus for the first time rather than Anthropic writing anything. Charting it would scale the axis past every real day since. Its date is not excluded, only those first captures: the same day carries genuine edits and they are here.
What was switched on when
352 readings352 gate readings across 19 releases, 2.1.234 to 2.1.269.
- unknown 161
- off 103
- on-in-code 38
- on-in-practice 37
- graduated 13
Claude Code reads its feature gates from a service at startup, so what a build does is not only what its code does. This is the reading taken by each release's own run, in the state the gate was actually in.
Account delta At most 1 gate in any of these releases was set differently for this account than for an anonymous one. The gate service answers per account, so a reading is this account's reading; where it differs from the anonymous baseline, the difference is counted here rather than hidden.
Range 19 releases. There is a second reading in the database, one ambient snapshot taken on one evening and attached to every older release at once, and it is deliberately not drawn: one evening's reading presented as the history of eighty-nine past releases would be a timeline of something that never happened. What a gate reading is and is not.