Nine of the ten paid AI coding plans reviewed changed a published usage limit at least once in 2026. This page is the record: 38 dated changes between February 13 and August 29, 2026, each with what the limit was, what it became, when it was announced, when it took effect, and two things that are usually collapsed into one — how the vendor described the change, and the direction a subscriber already on that plan experienced on the day it landed.
Those last two are separate columns, and they are never merged. In fourteen rows they disagree. That is the reason this table exists in the shape it does: a writer arguing that AI plan limits are unstable, and a writer arguing that vendors are candid about reductions, can both cite it, cell by cell, without either of them having to trust our verdict.
The most-covered change of the year is in here as two of the 38 rows. It is not the most interesting one. The sharpest row is a vendor that cut a quota without announcing it, one day after the deadline that decided whether you kept your old terms — and the second sharpest is a limit restoration that a vendor announced, then quietly did not perform for four weeks.
- 01Thirty-eight dated changes, nine vendors, seven months.Every row is an event dated on or before August 29, 2026, carrying its own source URL. Anthropic accounts for 11 rows, OpenAI 8, Z.ai 5, GitHub 5, Cursor 3, Alibaba 3, and Windsurf, Google and Moonshot one each. A tenth vendor, Amazon Kiro, was reviewed and is reported with zero qualifying rows rather than dropped.
- 02In 14 rows the vendor’s framing and the effective direction disagree.Thirteen are framed as an increase, an improvement, or neutrally, and were effectively a decrease for someone already on the plan. The fourteenth, Kimi’s, is framed as an increase and was effectively no change, because the vendor’s own help centre said the improved plan was not purchasable on the day it was announced.
- 03Two changes were never announced at all.Z.ai added a weekly ceiling to the GLM Coding Plan as an unannounced edit to a documentation page, landing after the grandfathering deadline that determined whether a subscriber kept the old terms. Anthropic raised the weekday peak-hour burn rate on Claude Code before saying so, and confirmed it only after users measured it.
- 04Four vendors publish no absolute figure for at least one paid tier.Anthropic, Cursor, Google and Windsurf all describe a paid limit without stating a number. Those rows are therefore expressed as an index, a multiple or a direction — never as a count. Cursor’s August 2026 rows are marked Not established for exactly this reason: the company published neither the new limits nor the new tariff.
- 05Two rows are scheduled, not done — read them as announcements.Anthropic’s permanent weekly-limit change takes effect September 14, 2026 and GitHub’s Business and Enterprise promotional credits end September 1, 2026. Both were announced on or before the as-of date and had not happened when this ledger was compiled. They are marked scheduled in the table and must not be cited as events that occurred.
01 — February 2026The sharpest row is a documentation edit. #
On February 11, 2026, Z.ai announced price rises for its GLM Coding Plan. The announcement covered prices and nothing else. It carried a grandfathering deadline of February 12: subscribers who had bought before that date with auto-renew enabled would keep their existing terms.
Between February 13 and February 16 — after the deadline had passed — the plan’s quota structure changed. A weekly ceiling appeared where none had existed, at roughly 400, 2,000 and 8,000 prompts for Lite, Pro and Max, alongside five-hour allowances of roughly 80, 400 and 1,600 prompts. Those after figures are not disputed: Z.ai’s own later notice tabulates them as the “Legacy Plan V2” reference. The change was never announced. It arrived as an edit to the docs.z.ai/devpack/overview
documentation page.
The before figures are weaker evidence, and the row says so. Z.ai does not restate the pre-edit five-hour allowance on any current page. A third-party reconstruction drawn from Internet Archive snapshots of the same documentation puts it near 120, 600 and 2,400 prompts, which would make the five-hour cut about a third — we could not confirm that first-hand, so the ledger does not carry the magnitude as a finding. The row does not need it. A weekly ceiling that did not previously exist is unambiguously a reduction in what an existing subscriber could do, and Z.ai’s own notice establishes that the grandfathered V1 cohort was the cohort with no weekly cap.
not been written down. The consequential part of the change was not in the announcement at all — it was in a documentation diff that landed after the decision window closed. This has had almost no English-language coverage.
The sequence did not stop there. On April 21 Z.ai announced that legacy plans “will no longer remain as a supported subscription option going forward”; on April 30 it cancelled auto-renewal on every eligible Legacy Plan V1 subscription, with two complimentary months and a 50% migration discount as compensation. The terms whose permanence was the reason to meet the February deadline were withdrawn ten weeks later. Those are rows 31 and 32. Our standing read of whether that plan is worth buying sits in the GLM Coding Plan value analysis; this page only records the dates the terms moved.
02 — July–August 2026A restoration that was announced, then postponed. #
The second finding is a category the market’s coverage does not have: an announced change that did not happen, discoverable only because the same person announced it twice.
On July 12, 2026, during GPT-5.6 Sol launch week, OpenAI temporarily removed the rolling five-hour usage limit for Plus, Business and Pro across ChatGPT Work and Codex, leaving weekly limits untouched. That is row 17, and it is an increase. On July 29, Codex and ChatGPT lead Tibo Sottiaux said the five-hour limit would come back the following day, alongside a full usage reset. That is row 18.
“Tomorrow we will bring back the 5h limit for Plus accounts across ChatGPT Work and Codex. I had mentioned this a while ago, but then postponed it.”— Tibo Sottiaux, OpenAI, on X, August 25, 2026
That sentence, posted on August 25, is direct evidence that the July 30 restoration did not hold for Plus accounts. Both dates are in the ledger, as they should be — the announcement was made and is a matter of record. Row 18’s effective direction is Not established, because whether the limit actually returned on July 30 for Plus is not something the public record settles. Row 19 carries the restoration that did land, on August 26, with Pro $100 and Pro $200 subscriptions explicitly exempted.
The practical point for anyone budgeting: an announced limit change is not the same object as an enforced one, and the gap between them was four weeks here. The launch-week pool structure that makes these rows matter across surfaces — ChatGPT Work, Codex and Workspace Agents drawing on one shared allowance — is covered in our GPT-5.6 launch-week usage pool breakdown.
03 — MethodWhat was collected, and what was left out. #
The exclusions do more work here than the inclusions. Most coverage of “AI plan changes” mixes prices, seat caps, new tiers and overage behaviour into one story, which makes the resulting numbers uncountable. This ledger holds exactly one object — a dated change to a usage ceiling on a paid plan — and names everything it turned away.
What was collected. Every dated change, in calendar 2026, to a published or enforced usage limit on a paid subscription plan for an AI coding tool. A change qualifies if it moved a ceiling, moved the rate at which a ceiling is approached, changed the unit the ceiling is denominated in, or withdrew a plan’s ability to keep its existing ceiling.
Vendors reviewed (10). Anthropic (Claude Code), OpenAI (Codex and the ChatGPT plans it draws on), Cursor, GitHub Copilot, Windsurf / Devin Desktop, Google (Gemini CLI, Gemini Code Assist, Antigravity), Z.ai (GLM Coding Plan), Alibaba Cloud (Qwen Coding Plan and Token Plan), Moonshot AI (Kimi Code) and Amazon Kiro.
Sources. Vendor-first: the vendor’s own changelog, documentation, help centre, pricing page or announcement post wherever one exists. Where a vendor announced a change only on X, only by email to subscribers, or only as an unannounced documentation edit, the row cites the best available secondary record and says so in the row itself. Every row carries its own source; there is no page-wide sourcing claim.
Dates. Sources were fetched on August 30, 2026. The data’s as-of date is August 29, 2026 — no event dated after August 29 appears as a row. Two rows are changes announced on or before the as-of date that take effect after it; they are marked scheduled and must not be read as having happened.
N = 38 rows across 9 vendors. A tenth vendor, Amazon Kiro, was reviewed and is reported with zero qualifying rows rather than dropped. Where one announcement changed one thing in opposite directions for two populations, the row is counted once under the direction most subscribers experienced and its second direction is named in the row — so counting directions rather than rows gives more entries than 38.
What was excluded, and why. Eight classes. Price changes — Windsurf’s Pro rise, Z.ai’s February and April rises, Google’s Ultra cut: these change what you pay, not what you get. API rate limits — Anthropic’s May 6 Opus API increase and every per-token price move are API-side, not subscription-side. New tiers — OpenAI’s $100 Pro, GitHub’s Max, Kiro’s Pro Max, Windsurf’s Max, Cursor’s Start: a new plan adds an option, it does not move an existing subscriber’s ceiling (where a launch also changed an existing plan’s allowance, that part is a row). Seat and access limits — ChatGPT Business’s 200-seat cap, Copilot’s sign-up s, Kimi’s subscription : these limit who may buy, not how much a subscriber may use. Behaviour at the ceiling — Kimi’s Extra Usage, Kiro’s add-on credits and overage caps, Copilot’s Auto-only fallback: what happens when you hit the wall is a different object, covered by our spend-cap exhaustion census. Restatements — Cursor’s July 21 repetition of its July 16 doubling is not a second doubling. Effective-capacity changes with no change to the limit — tabulated separately below rather than mixed into the ledger. And pre-2026 events — Anthropic’s introduction of weekly limits in August 2025 is the baseline this ledger measures from, not a row in it.
Known limitations. Four of the nine vendors with rows publish no absolute figure for at least one paid tier’s limit (Anthropic, Cursor, Google, Windsurf), so those rows are expressed as an index, a multiple or a direction rather than a count. Six rows are marked Not comparable (unit change) because the vendor changed the unit and published no conversion. Two Anthropic extension dates are known to within a few days rather than to the day and are stated as ranges; the March promotion’s end date is given as March 27 or 28 because sources conflict. Several rows rest on a single record — an X post, an email forwarded to a forum, or a third-party reconstruction from archived documentation — and each such row says so. Row 30, the Z.ai silent cut, is the weakest-sourced row for its before values and carries an explicit caveat. Where a vendor made no characterisation at all, the framing cell says so rather than inferring one. Finally, this is a ledger of what was announced or enforced, not a measurement of what any account actually received: we did not instrument accounts on these plans.
04 — The LedgerThirty-eight dated changes, nine vendors. #
Two columns carry the weight. Vendor’s framing is the vendor’s own characterisation of the change, quoted where a quotation exists. Effective direction is the direction experienced by someone already subscribed on the day the change landed. They are deliberately separate cells and are never merged into a verdict. The effective-direction column uses a fixed vocabulary: Increase
, Decrease
, Decrease (scheduled)
, Neutral
, Not comparable (unit change)
, No change in practice
and Not established
.
One row concerns the most-reported change of the year. Anthropic announced on August 29 that it would permanently raise standard Claude Code weekly limits by 25% from September 14, and stated in its own follow-up that against the level in force that day this “works out to a 17% reduction in weekly limits on Claude Code.” The 17% is Anthropic’s figure, not ours. The arithmetic, the communications sequence and the correction of the coverage that got it wrong belong to our dedicated post on the September 14 change. Here it is rows 10 and 11 of 38, and it is interesting only comparatively.
| # | Vendor and plan | Before | After | Announced | Effective | Vendor’s framing | Effective direction | Source |
|---|---|---|---|---|---|---|---|---|
| Group A — Anthropic, Claude Code (11 rows) | ||||||||
| 1 | Anthropic — Claude Code plus Claude web, desktop, mobile, Cowork, Excel and PowerPoint. Free, Pro, Max 5x, Max 20x, Team. Enterprise excluded. | Standard five-hour limits at all hours. | 2× the five-hour limit outside weekday 08:00–14:00 ET and all day at weekends. Bonus usage did not count toward weekly caps. | 2026-03-12/13, @claudeai on X | 2026-03-13 to 2026-03-27 or 2026-03-28 — sources conflict | Increase — “a small thank you to everyone using Claude” | Increase | |
MacRumorsAnthropicAnthropicScriptXenoapidogScriptXenoScriptXenoOur July 20 coverageBleepingComputerBleepingComputerOpenAI Help CenterTheNextWebOpenAI Help CenterTheNextWebofox.aiMorphCodex RunwayexplainxStartup FortuneexplainxKiloGitHub changelogGitHub blogGitHub changelogGitHub blogGitHub docsWindsurf blogOur migration guidedocs.z.ai/devpack/overview
page, after the 2026-02-12 grandfathering deadline had passed.Z.ai docsZ.ai docsZ.ai docsZ.ai docsZ.aiAlibaba CloudAlibaba Cloudsk-sp-
API key and the coding.dashscope.aliyuncs.com
base URL.Alibaba CloudKimi Help CenterKiro changelogCounting rule. A row with two populations is counted once, under the direction most subscribers experienced; its second direction is named in the row. On that basis the 38 rows break down as 10 increases, 13 decreases, 2 scheduled decreases, 5 neutral continuations, 6 marked Not comparable (unit change)
, one No change in practice
and one Not established
. Three rows carry a second direction for a second population or component — rows 9, 33 and 37 — which is why counting every direction rather than every row gives a larger total.
05 — Framing vs EffectFourteen rows where the two columns point opposite ways. #
Rows 2, 9, 11, 13, 15, 19, 22, 26, 27, 29, 31, 32, 36 and 38 are the fourteen where the vendor’s characterisation and the effective direction disagree. Thirteen of them share a shape: framed as an increase, an improvement, or neutrally — effectively a decrease for someone already on the plan. The fourteenth, row 38, is different in kind: Kimi’s change was framed as an increase and produced no change in practice, because Moonshot’s own help centre said on the same day that the improved plan was not purchasable.
Row 30 is counted separately and is not one of the fourteen. Z.ai never characterised its February change at all, so there is no framing to disagree with. A missing framing is a different finding from a mismatched one, and collapsing the two would inflate the count by one.
Framed as an increase
A change announced as more capacity that removed capacity for at least one population. Row 9 gave Max and Team Premium a standing Fable 5 inclusion at up to half the same weekly cap, and took included access away from Pro and Team Standard.
Framed as neutral or as an upgrade
No direction claimed. Anthropic “adjusting” a session limit, Alibaba’s “due to a product upgrade”, a promotion ending on a date stated at launch, a legacy plan quietly becoming terminal. Neutral wording, downward effect.
Framed as a scheduled end
The vendor described a promotional window closing or a d limit returning — accurate, and still a reduction against the level the subscriber had been living with. GitHub’s September 1 reversion is the clean example.
Framed as an increase, effectively nothing
Moonshot announced the removal of Kimi’s monthly total cap on the same day its own help centre said the new plans could not be bought. Six weeks later the live docs still described the cap the announcement was meant to remove.
The one row where a vendor stated the direction plainly is 23. GitHub wrote “We are tightening usage limits for individual plans” and, in the same post, “We know these changes are disruptive, and we want to be clear about why we’re making them and how they will affect you.” Framing and effect agree, and the row is in the ledger as a decrease with no gap between the two columns. That row is the reason the framing column is worth keeping: without it, the column would just be a second decrease column with extra steps.
06 — The PatternNot up, not down — boosted, extended, then settled lower. #
Read the 38 rows in date order and the dominant 2026 pattern is not “limits went up” or “limits went down”. It is a temporary increase that expires, gets extended repeatedly, and then settles below the level users had grown used to. Anthropic ran five extensions of one promotion. OpenAI ran a 10×-to-5× promotional revert and two removals-and-restorations of the same five-hour limit. Z.ai ran a 1.5× boost with a hard end date two days after this ledger closes.
That is where nearly every framing-versus-effect disagreement comes from. The subscriber’s reference point is the boosted number they have been working against for months. The vendor’s reference point is the baseline before the promotion started. Both parties can describe the same move honestly and arrive at opposite adjectives. The chart below is the clearest single case, drawn from rows 5 through 11.
Held against the index, both of Anthropic’s statements are exact. 125 is 25% above 100 and 17% below 150. The disagreement is not arithmetic; it is which number sits in the denominator, and after fifteen weeks at 150 the subscriber’s denominator is 150. The instance in full — the arithmetic, the deleted-and-reposted thread, the Help Center discrepancy — is in our September 14 post.
07 — Negative FindingsThe rows that say nothing moved. #
A ledger that lists only vendors who moved overstates how unstable the market is. Two findings in this dataset are negatives, and both are load-bearing.
The first is Group J. Amazon Kiro was reviewed and has zero qualifying rows. Its 2026 changelog is not empty — a new Pro Max tier in June, a free student allocation in March, add-on credits replacing the overage model in July, enterprise overage caps the day after — but none of that changed the included allowance on a paid tier that already existed. Fifty, one thousand, two thousand, five thousand and ten thousand credits per month are the same figures at the end of August as at the start of the year. The row is reported as “no qualifying row found”, not as “no change occurred”, because those are different claims. Kiro’s wider product shape is in our Kiro guide.
The second is Cursor, marked Not established. Cursor raised included usage limits and changed how Auto bills, in the same email, on the same day, and published neither the new limit figures nor the new per-model tariff — not on the blog, not in the changelog, not in the docs, which still listed Auto at a flat rate afterwards. Rows 21 and 22 therefore record an increase of unstated magnitude and a decrease in effective capacity, and the net is Not established
. A reconstructed number would be more satisfying and less true. “The vendor did not say” is the citable finding.
Amazon Kiro
Reviewed across the full 2026 changelog and the current billing docs. New tiers and overage mechanics changed; the included allowance on every pre-existing paid tier did not.
Cursor, August 2026
An increase and a tariff change announced together by email, with no figure published for either. Rows 21 and 22 stand as Not established rather than being netted into a guess.
Vendors publishing no figure
Anthropic, Cursor, Google and Windsurf each describe at least one paid tier's limit without stating a number. Those rows are indexes, multiples or directions — never counts.
Eight further candidates were investigated and did not become rows. Listing them is part of the method, not an appendix: it shows where the boundary of the class sits and what evidence was judged insufficient.
| Candidate | What the record shows | Why it is not a row |
|---|---|---|
| OpenAI silently extending the five-hour limit to Team and Business workspaces, on or about 2026-08-26 | One poster on OpenAI’s own developer community forum states the limit was enabled on Team and Business workspace accounts without being disclosed in the announcement. | A single user report on a support forum, with no vendor confirmation and no second independent account. If true it would be the ledger’s third silent change and the most consequential — which is why it needs better evidence, not less. |
| Windsurf removing the daily limit for Max users, April 2026 | Surfaced only in a search-result synthesis. No vendor page, changelog entry, blog post or dated report located. | No primary source and no date. |
| Google merging Antigravity’s Gemini Flash and Gemini Pro rate-limit pools, May 2026 | A third-party page cites “Google’s May 2026 plan-change announcement” for a merge of two model pools into one shared rate limit, with non-Gemini models on a separate fixed limit. | Month precision only. The pool merge is a genuine limit change and probably belongs here; this asset’s contract is that every row is a dated event, and no day was established. |
| The exact date Anthropic first extended the +50% weekly promotion | Two sources place the first extension in mid-July 2026; neither gives the day. | Partially verified — the extension happened and is row 6, which states the date as “on or about 2026-07-13” rather than claiming a day the record does not support. |
| The February 2026 Claude Code prompt-caching bug and the rate-limit reset that followed | Referred to in March 2026 reporting as having happened “less than a month ago”, with no date given. | Rejected on two grounds: no date, and it is an effective-capacity event rather than a limit change. It appears in the adjacent-events table below instead. |
| Z.ai’s pre-February five-hour prompt figures | Approximately 120, 600 and 2,400 prompts per five hours, reconstructed by a third party from Internet Archive snapshots of Z.ai’s documentation. | Conditionally admitted as context in row 30 with the reconstruction named, and excluded as a magnitude claim: the ledger does not state the size of the five-hour cut. |
| The day-level sequence of Z.ai’s new weekly cap in February 2026 | A community record places the new weekly cap at 4× the five-hour allowance on 2026-02-14 and 5× on 2026-02-16. | Single community source. Row 30 therefore gives the window as 2026-02-13 to 2026-02-16 and only the final figures, which Z.ai’s own notice confirms. |
| Amazon Kiro’s Free-tier allowance history across 2026 | Kiro’s changelog and billing docs give the current figure of 50 credits but no 2026 change history for the Free tier specifically. | Not a row. Absence of a recorded change is what Group J reports, and it is reported as “no qualifying row found” rather than as “no change occurred”. |
08 — StructureHow many ceilings you are actually under. #
A single-window mental model — “I get X per month” — is wrong on almost every plan in this ledger. Counted from the vendor documentation cited above, seven of the nine vendors with rows meter a coding plan against two or more simultaneous windows, and the window that stops you is whichever fills first. This is derived entirely from sources already cited per row, and it is the cheapest available citation for anyone arguing that these plans are hard to budget against.
| Vendor | Windows enforced at once | Publishes an absolute figure? |
|---|---|---|
| Anthropic — Claude Code | Rolling five-hour + weekly | No — no token or prompt figure for any Claude Code limit |
| OpenAI — Codex and ChatGPT Work | Rolling five-hour + weekly, on a shared credit pool across surfaces | Yes — a published credit rate card |
| Cursor | Included usage per billing cycle, across two separate pools | Partly — dollar figures for the third-party pool, none for the Cursor Models pool |
| GitHub Copilot | Calendar month, resetting at 00:00 UTC on the first, with no rollover | Yes — credits per user per month |
| Windsurf / Devin Desktop | Daily + weekly | No — quotas described qualitatively rather than numerically |
| Google — Antigravity | Five-hour refresh under a weekly compute ceiling | No — quotas described as “meaningful”, “high, generous” and “the highest, most generous” |
| Z.ai — GLM Coding Plan | Five-hour + weekly | Yes — credit figures per tier, plus peak and off-peak rates |
| Alibaba Cloud — Token Plan | Rolling five-hour + rolling seven-day | Yes for credits — but no published token value per credit, per model |
| Moonshot AI — Kimi Code | Five-hour + weekly + the membership’s monthly total | Yes — credits refreshing every 7 days from the subscription date, with no carry-over |
One class of event was deliberately excluded from the ledger and is shown here instead, because it is the class readers most often mistake for a limit change: something moved how far an allowance goes without moving the allowance. Promoting the exclusion to a visible table is how the methodology stays auditable rather than merely asserted.
| Date | Vendor | What moved | What did not move | Source |
|---|---|---|---|---|
| February 2026, day not established | Anthropic | A prompt-caching bug caused Claude Code usage to drain faster than expected; Anthropic reset rate limits afterwards. | The published limits. | |
Codex RunwayZ.ai## 09 — ImplicationsBudgeting against a ceiling that moves. Thirty-eight changes across seven months is roughly one every five days somewhere in the market, which is a rate, not a crisis. The useful reading is not that plans are unreliable but that the reference point in a subscriber’s head decays. A boosted allowance becomes the normal allowance within weeks, and every subsequent change is then measured against a number the vendor never promised.
Write down which number your plan actually promises
Rows 5 through 11 sat at a promotional level for fifteen weeks. Teams that budgeted against 150 and not 100 experienced September 14 as a cut. Note the promotional end date next to the allowance, in the same place.
Assume more than one ceiling
Seven of the nine vendors with rows enforce two or more windows at once, and one enforces three. A monthly figure tells you nothing about whether a Tuesday afternoon will stop.
A unit change is not a limit change you can compare
Six rows are marked Not comparable because a vendor moved from requests to tokens or credits to quotas without publishing a conversion. Re-measure your own consumption after any unit change; the vendor's before-and-after does not net out.
Two of the 38 changes were never announced
One arrived as a documentation edit after a grandfathering deadline; one was confirmed only after users measured it. If a plan matters to your delivery schedule, watch its docs page, not only its changelog.
Two neighbouring questions have their own pages, deliberately fenced off from this one. What happens at the ceiling — hard stop, error, degrade, auto-overage — is a different object, censused in our spend-cap exhaustion behaviour census. What you pay, as opposed to what you get, is covered in our comparison of subscription, token-plan and credit pricing models. Every price change in 2026 is excluded from this ledger for that reason. The sibling asset in the same shape, tracking a different event class, is our ledger of model supply cutoffs. And if you want a second set of senior eyes on which coding plans your team should be standardising on, and what a moving ceiling does to your delivery estimates, that is the kind of question our AI transformation engagements open with.
10 — ConclusionTwo columns, kept apart. #
Thirty-eight dated rows, fourteen disagreements, and one vendor that changed nothing.
The single design decision that makes this table worth citing is that it never reaches a verdict. The vendor’s framing and the effective direction sit in adjacent cells, each carrying one claim, and in fourteen rows they contradict each other in public. Nobody has to trust our reading to use the data — they can quote the cell that supports the argument they are making, and the neighbouring cell will still be there.
The strongest row is not the one that got the coverage. Z.ai cut a quota as a documentation edit, after the deadline that decided whether you kept the old terms. OpenAI announced a limit restoration and then, four weeks later, revealed it had been postponed. Both are more consequential to a subscriber than any single percentage, and both were nearly invisible in the coverage.
The negatives keep the rest honest. Amazon Kiro was reviewed and changed nothing about an existing paid tier’s allowance all year. Cursor’s two August rows say Not established
, because Cursor published no figure and reconstructing one would be a guess wearing a number’s clothes. Rows 11 and 27 are marked scheduled and had not happened when this page was compiled. We will append rows as dated changes land, move the scheduled ones when their dates arrive, and re-date the page rather than replace it.