三家 AI 供应商在两周内调整了使用限制：如果你为 AI 付费，这意味着什么

Between August 25 and September 10, 2026, Anthropic, OpenAI, and GitHub Copilot each changed the usage limits on their paid AI plans. Not one announcement said "we are cutting your access," and some presented the change as an increase. The math told a different story

When three major AI vendors reshape their usage limits within the same two weeks, it gets much harder to treat any one of them as an isolated decision. A pattern is showing itself.

It isn't a coincidence of timing. After roughly eighteen months of promotional generosity, the AI subscription market is finding its permanent price floor. The promotions kept users engaged while competition was at its fiercest. Now they are expiring, settling, or being recalibrated, because the compute cost of frontier models demands more careful rationing. A subscriber's access either holds or it doesn't, and the way these adjustments are communicated often makes it hard to tell which.

Below is what actually changed at each of the three vendors, why it matters, and what to track if your work depends on AI tools you pay for every month.

What Anthropic changed: a "25% increase" that is a 17% cut

On August 29, 2026, Anthropic announced a permanent 25% increase to standard weekly usage limits for Claude Code, starting September 14. The announcement called it an increase, and that description is technically correct.

The context is what matters. Since May 13, 2026, Claude Code subscribers had been working with a weekly limit 50% above the pre-promotional baseline. That boost was prolonged through five successive extensions, each of which bought a new end date without committing to a permanent level. Subscribers spent fifteen weeks at the higher ceiling. Many teams built workflows around it, and some even staffed against it.

When September 14 arrives, the weekly limit becomes 125% of the original pre-promotional baseline. Measured against that old number, it's +25%. Measured against the level in force since May, it's -17%. Anthropic's own communications confirmed both figures: "we're permanently raising standard weekly limits in Claude Code by 25%" and, in a follow-up, "this works out to a 17% reduction in weekly limits on Claude Code" against the level in force on the day of the announcement.

Both statements are arithmetically exact. The disagreement isn't over the numbers but over which one sits in the denominator. For the vendor, the reference point is the pre-promotion baseline from six months ago. For the subscriber, it's the ceiling they've been hitting for fifteen weeks. The math is the same, and the adjectives are opposite.

The index in practice: Anthropic publishes no absolute figure for any Claude Code weekly limit on any paid tier: not the baseline, not the boosted level, and not the permanent level from September 14. The only verifiable figures are relative: 100 (baseline), 150 (promotion), and 125 (permanent from September 14). If you want to know how many tokens, prompts, or requests your plan really allows, Anthropic's documentation doesn't contain that number.

What OpenAI changed: the 5-hour limit that kept disappearing

OpenAI's two-week window holds two separate changes. They're worth pulling apart, because each one affects a different set of subscribers in a different way.

On August 25, Tibo Sottiaux, the Codex and ChatGPT lead at OpenAI, announced that the rolling five-hour usage limit would return for Plus accounts the next day. His exact wording: "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." That sentence carries a quiet admission. The limit had first been announced as returning on July 30, then announced again on August 25, and Plus subscribers went without a five-hour cap for the four weeks in between.

The sequence behind it is documented. On July 12, OpenAI temporarily removed the rolling five-hour limit altogether during GPT-5.6 Sol launch week. On July 29, it announced the limit would be back on July 30, but it didn't return for Plus subscribers that day. The August 25 announcement was the second attempt at the same restoration. In total, the limit was gone for more than six weeks. That absence was partly unintentional and partly postponed, and it was communicated across at least three separate posts, none of which said "we postponed this last time."

The second OpenAI event is sharper. On September 9, OpenAI cut GPT-6 Astra usage limits by up to four times for heavy users, and on September 10 it confirmed that new $200 Pro plan subscriptions were being paused to protect service quality for existing subscribers. OpenAI is also testing a pay-to-reset option: $5 to $8 to instantly refill a hit limit on the Plus plan, rising to $50 to $80 on Pro. Since June, OpenAI had offered free banked resets. That same relief now comes with a price tag.

What GitHub Copilot changed: the promotional window closes

Of the three, GitHub Copilot's September 1 change was the most clearly disclosed. When Copilot Business and Enterprise customers moved to usage-based billing on June 1, 2026, they received promotional included credits: 3,000 per user per month for Business and 7,000 for Enterprise. GitHub's own billing documentation gave the promotional window as June 1 to September 1.

On September 1, the promotional credits ended and the standard included credits took over: 1,900 credits per user per month for Business (a 37% reduction from the promotional level), and 3,900 for Enterprise (a 44% reduction). A ten-person Business team that had been working with 30,000 credits a month now has a September floor of 19,000. If that team didn't adjust its workflow volume during the promotional period, it now faces a gap.

GitHub's transparency here deserves explicit credit. The end date sat in the billing documentation from the start, the standard credits were the committed baseline, and the promotional level was always labeled as promotional. The September 1 reduction was disclosed in advance and spelled out in the docs, and GitHub's announcement framing matched the effective direction. Most of the other changes in the same window weren't communicated with that kind of clarity.

The broader pattern these three weeks reveal

Digital Applied has published a detailed ledger that tracks 38 dated changes to paid AI coding plan limits across nine vendors, from February to late August 2026. Its findings are worth knowing because they describe the market as a whole, not just three individual events.

"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." Digital Applied, AI Coding Plan Limit Change Ledger, August 29, 2026 In 14 of the 38 rows, the vendor's framing and the subscriber's effective direction point opposite ways.Thirteen of those 14 were framed as increases, improvements, or neutral changes, while the subscriber on the day of the change experienced a decrease.

The last one was framed as an increase that produced no change at all, because the vendor's own help center said the improved plan couldn't be purchased on the announced date.

Two changes were never announced by the vendor at any point. One arrived as a documentation edit after the grandfathering deadline that decided whether a subscriber kept their old terms. The other was confirmed only after users measured a faster-than-expected rate of limit consumption. If a plan matters to your delivery schedule, watching its documentation page gives earlier warning than the changelog or official announcements.

Four vendors in the ledger publish no absolute figure for at least one paid tier's limit: Anthropic, Cursor, Google, and Windsurf each describe a paid plan's ceiling without stating a number. Their subscribers can't verify their allowance from any vendor-published source. They can measure their consumption and infer, but they can't compare what the pricing page says with what they experience in production.

The multi-window problem

Seven of the nine vendors in the ledger enforce two or more simultaneous usage windows, and one enforces three. For budgeting, that matters in ways a monthly headline figure hides.

A plan whose homepage promises "X units per month" may also enforce a five-hour rolling window and a weekly ceiling. Whichever of the three fills first is the one that stops you, no matter how much remains in the others. A team running a heavy processing job on a Tuesday afternoon can exhaust a five-hour window while 80% of its weekly allowance is still unspent. The job stops, and the weekly budget doesn't help.

Anthropic (Claude Code): A rolling five-hour window plus a weekly ceiling.No absolute figure is published for either. OpenAI (Codex and ChatGPT Work): A rolling five-hour window plus a weekly ceiling, drawn from a shared credit pool across several surfaces at once. GitHub Copilot (Business/Enterprise): A calendar-month reset with no rollover, plus model-specific multipliers that set how fast the monthly pool drains. Google (Antigravity): A five-hour refresh under a weekly compute ceiling.No absolute figure is published at any tier.Teams that budget against the monthly headline number and skip the sub-limits are planning around a figure that isn't their real constraint.

What counts is whichever window fills first for the team's actual usage pattern.

What this means if you pay for AI

The situation in September 2026 comes down to three things. Promotional generosity is ending across the major AI platforms at the same time. Compute costs are being handled through tighter rationing rather than higher subscription prices. And the means of finding out what you actually have available are often opaque, whether by design or by inattention.

The pay-to-reset option OpenAI is testing is a signal worth reading closely. It turns a hit rate limit from a hard stop into a purchasable override, which changes the pricing model in a direction that's predictable for the vendor and unpredictable for anyone who budgeted around a flat monthly fee. If the pattern holds, a flat subscription buys baseline access, peak-hour or high-volume usage is metered on top of it, and the effective cost of an AI-intensive workflow rises with usage intensity instead of staying constant. That's a different economic model from the one most teams priced against when they chose an AI plan in 2025.

"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." Digital Applied, AI Coding Plan Limit Change Ledger, August 29, 2026

Here are four practical responses to what's changing:

Record your baseline, not the promotion. If your plan has been running a temporary increase, write down its stated end date and the baseline it will settle at, and budget against that number. Fifteen weeks at a boosted ceiling becomes your mental reference point whether or not it's the committed one.

Know which window stops you first. Work out whether your real usage is more constrained by the five-hour rolling limit, the weekly ceiling, or the monthly total.The answer differs by workflow: a batch job is more likely to hit the five-hour limit, while a distributed team's combined usage is more likely to hit the weekly or monthly one. Watch the documentation, not the changelog. Two of the 38 changes in the Digital Applied ledger were never announced, and both surfaced as documentation changes.

If a plan's ceiling matters to your production schedule, a diff of the relevant documentation page warns you earlier than any announcement channel. Do not treat an announced change as an enforced one. OpenAI announced the five-hour limit's return on July 29, and it reached Plus on August 26.Nearly four weeks separated those dates, and the subscription price never changed in between.Measure your actual limits against production usage, not against the latest announcement.

The structural shift underneath all of this

The promotional era of AI subscriptions did its job. It built user bases at scale, established habits of daily use, and created the market that vendors are now trying to monetise more precisely. What's happening across the major platforms in late 2026 is the natural next step: limits tighten, usage intensity gets metered, and the economic structure shifts toward something that better reflects the compute cost of the models being used.

GPT-6 Astra limits being cut four times for heavy users is an extreme case, but it reflects the underlying pressure. The most capable frontier models are genuinely expensive to run at scale, and a flat $20-$200 monthly subscription can't cover that cost once usage becomes intensive enough. Vendors are working through the gap between what they charged during the growth phase and what the compute economics require, in a market now mature enough to accept more granular pricing.

For subscribers, the right frame is neither "they are cutting my access" nor "they are raising their prices." It's that the steady-state pricing model for AI tools is being set in real time and in public, through a series of adjustments that are individually small and collectively significant. The promotional baseline is on its way out. The permanent one is being established right now.

The team that understands its plan's multi-window structure, tracks the documentation rather than the announcements, and budgets against the committed baseline instead of the promotional level is the team that won't be caught off guard when the next adjustment lands. That adjustment isn't a question of if. It's a question of when, and for which window.

Enjoyed this read? Be sure to follow me for more stories that make people feel things. You can also connect with me on LinkedIn or Telegram @kawsarlog for more updates on my work.

I do freelance scraping and data automation projects. Feel free to contact me on https://www.linkedin.com/in/kawsarlog/ or email at hi@kawsarlog.com