Every Time an AI Coding Plan Changed Its Limits in 2026

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.

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.

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.

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.

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.

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.

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.

Counting 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.

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.

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.

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.

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.

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.

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.

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.

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.

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.