AI coding limits stopped being numbers

In July 2025, 7 of 9 AI coding plans stated their main limit as a number. By November 2026, 3 did. We tracked 62 limit changes across six vendors.

Cover for AI coding limits stopped being numbers: 7 of 9 tracked plans stated their main limit as a number in July 2025 and 3 of 9 in November 2026, shown as nine plan cards whose vague limits are blurred, leaving ChatGPT Plus (160 messages every 3 hours), Codex in Plus (5–45 to 350–3,000 messages per 5 hours) and Copilot Pro (1,000 + 500 credits) legible; usage meters show Claude Max 20x giving 6 times Pro a week and Max 5x 3.5 times; 62 announced limit changes across 6 vendors, January 2024 to October 2026.
On this page

In April 2024 Anthropic told Claude Pro subscribers they could expect to send "around 45 messages every 5 hours". Today Claude's pricing page promises Pro "more usage than Free", every five hours and every week. Its top plan, Ultra at $200 a month, offers "10x Pro usage every week" and no five-hour limit. That's not one plan. It's how the whole category moved: limits that used to be a count of messages or requests became windows, weekly caps, multiples of another plan and credits.

We tracked it. We read the archived versions of 19 help-center, pricing and docs pages from Anthropic, OpenAI, Cursor, GitHub, Google and Windsurf from January 2024 to 1 November 2026, plus the git history of GitHub's Copilot docs and Google's Gemini CLI. That gave us 62 announced changes to a published limit, each dated and quoted. Then we counted how people reacted: GitHub issues, Hacker News stories, Cursor's forum and downloads of a tool that meters your own usage. The methodology explains how, and what the archive misses.

Key findings

  1. 1
    The numbers went away. In July 2025, 7 of the 9 plans we tracked stated their main limit as a number on the vendor's own page. In November 2026, 3 do: ChatGPT Plus (160 messages every 3 hours), Copilot Pro (AI credits) and Codex in Plus (per-model estimates of messages per five hours).
  2. 2
    Limits change all the time. We counted 62 announced changes to published limits across six vendors: 5 in 2024, 26 in 2025 and 31 in the first ten months of 2026. Anthropic made 19, OpenAI 17 and GitHub 11.
  3. 3
    They didn't simply get tighter. Of 51 changes to existing plans, 21 tightened a limit, 18 loosened one and 10 were mixed. What changed reliably is how limits are stated: 76% of changes in the second half of 2025 published the new limit as a number, against 17% so far in the second half of 2026.
  4. 4
    "20x" was a per-session promise. By Anthropic's own weekly estimates, Max "20x" gave about 6 times Pro's usage a week and Max "5x" about 3.5 times: 240–480 and 140–280 hours of Sonnet 4, against Pro's 40–80.
  5. 5
    Every big plan now has a week. A short rolling window plus a weekly cap is now the shape at Anthropic, OpenAI's Codex, Google's Gemini app and Windsurf; Cursor's new plans have a weekly pool. GitHub alone still counts by the month.
  6. 6
    Users noticed the week. Claude Code GitHub issues with "weekly limit" in the title went from none before July 2025 to 286 by October 2026. Hacker News stories about AI coding-tool limits went from 15 in 2024 to 96 in 2025 and 327 in January–October 2026.
  7. 7
    Cuts and billing changes get the attention. The biggest Hacker News limit thread was GitHub's move to usage-based billing (767 points). Limit stories in the week of Anthropic's holiday 2x boost drew 12 points in all; the week it ended drew 31.

7 → 3

Of 9 tracked plans, those whose page states the main limit as a number, Jul 2025 → Nov 2026

Source: host0 analysis of vendor pages

6×

Claude Max 20x's weekly allowance vs Pro, by Anthropic's own estimates (Max 5x: 3.5×)

Source: Anthropic via TechCrunch

62

Announced limit changes across six vendors, Jan 2024 – Oct 2026 (31 in 2026)

Source: host0 changelog of archived pages

0 → 286

Claude Code issues with "weekly limit" in the title, before Jul 2025 → total by Oct 2026

Source: host0 count of GitHub issues

62 changes in 34 months#

A "change" here is an announced change to a published limit, to how a plan's limit is structured (a new window, credits instead of messages), to its top tier, or to whether the limit is disclosed at all. One-off counter resets don't count, and several rows from one announcement count once.

Announced limit changes per vendor

Changes to published plan limits, Jan 2024 – Oct 2026

Announced limit changes per vendor
LabelValue
Anthropic19 (2 · 8 · 9)
OpenAI17 (2 · 6 · 9)
GitHub11 (1 · 2 · 8)
Google7 (0 · 4 · 3)
Cursor5 (0 · 4 · 1)
Windsurf3 (0 · 2 · 1)
The notes split each total into 2024 · 2025 · 2026. All six vendors together: 5 changes in 2024, 26 in 2025 and 31 in January–October 2026. 2024 is under-counted: the archive holds few captures of that year's help-center articles.

Source: host0 changelog of archived vendor pages (changes.csv)

Anthropic changed Claude's limits most often. In 2024 it cut Pro's reset from eight hours to five and its estimate from "at least 100" messages to "around 45". In 2025 it launched Max, added Claude Code to Pro, then announced weekly limits on 28 July. In 2026 it tightened peak hours, doubled the five-hour window, ran and ended a weekly boost, and by 1 November had replaced Max 20x with Ultra. GitHub's eight changes in 2026 were mostly one long move from premium requests to usage-based AI credits.

It would be easy to read all this as limits getting stingier. Our count doesn't say that. We classified every change to an existing plan by a rule written before counting: fewer units per window, a new cap, a higher multiplier or work starting to count is "tightened"; the reverse is "loosened"; a change of unit with no published conversion is "mixed".

Which way the changes went

Changes to existing plans' limits and structure, Jan 2024 – Oct 2026 (n = 51)

Which way the changes went
LabelValue
Tightened21 (1 · 10 · 10)
Loosened18 (1 · 8 · 9)
Mixed10 (0 · 1 · 9)
Neutral2 (1 · 1 · 0)
Notes split each count into 2024 · 2025 · 2026. Tightened and loosened are about even. The highlighted mixed changes, where the unit itself changed, went from 1 in 2025 to 9 in 2026.

Source: host0 analysis of changes.csv

Dropping the changes we couldn't verify in our own archive (sourced only from news reports or a vendor announcement) gives 17 tightened, 14 loosened and 10 mixed: still about even. The robust trend is the mixed column, and the one in the next section.

The numbers disappeared#

For each of nine plans we checked how the vendor's own page stated its main limit in July 2025 and in November 2026: as an exact number per window, as an estimate (a range, "about", "average users"), only as a multiple of another plan ("5x Pro"), or not at all.

How each plan states its main limit

Vendor's own page, July 2025 → November 2026

  • Claude Pro: estimate (about 45 messages, 10–40 Claude Code prompts per 5 h) → none ("more usage than Free")
  • Claude's top plan: estimate (Max 20x: 200–800 prompts per 5 h) → relative (Ultra: 10x Pro a week)
  • ChatGPT Plus: exact (80 per 3 h) → exact (160 per 3 h)
  • Codex in Plus: none ("generous") → estimate (5–45 to 350–3,000 messages per 5 h, by model)
  • Cursor's entry plan: exact ($20 of usage) → relative (Premium: 4x Plus)
  • Copilot Pro: exact (300 premium requests) → exact (1,000 + 500 AI credits)
  • Windsurf Pro: exact (500 prompt credits) → none ("increased quotas")
  • Gemini app, AI Pro: none → relative (4x standard)
  • Gemini CLI, free: exact (1,000 requests a day) → none (moved to Antigravity CLI)
7 of 9 plans stated a number (exact or estimate) in July 2025, 3 of 9 in November 2026. Codex in Plus moved the other way, from no number to per-model estimates with ranges as wide as 15–160 messages. Dropping Codex and Gemini CLI gives 6 of 7 → 2 of 7.

Source: host0 analysis of archived and live vendor pages (disclosure_checkpoints.csv)

The pages that dropped numbers mostly replaced them with multiples. Cursor's plans are now sold inside the X Subscription, where Premium is "4x the usage" of Plus, and Plus is a weekly pool of unstated size. Google's Gemini app has "compute-based usage limits": AI Pro is "4x higher than standard limits", and standard isn't a number either. Windsurf, now Devin Desktop, sells "increased quotas" that refresh daily and weekly.

Even where a number survives, it's hedged. OpenAI's Codex page gives Plus users between 5 and 45 local messages per five hours on its flagship model, and adds: "These estimates are not fixed message limits". The same shift shows in the changes themselves.

Changes that published the new limit as a number

Share of announced changes whose new limit is an exact number or an estimate, by half-year

Changes that published the new limit as a number
LabelValue
2024 H1100% (n=3)
2024 H250% (n=2)
2025 H173% (n=15)
2025 H276% (n=17)
2026 H142% (n=26)
2026 H217% (n=12)
2026 H2 runs to October. The rest of each half is changes stated only as a multiple ("5x Pro") or with no number at all. 2024's samples are too small to read on their own.

Source: host0 analysis of changes.csv

"20x" was a per-session promise#

When Anthropic launched Max in April 2025, its help center offered "20x more usage than Pro". On 28 July it announced weekly limits from 28 August, and by the capture of 29 July the line read "20x more usage per session than Pro". The weekly estimates Anthropic gave TechCrunch, and later put in its own help center, show what that qualifier was worth.

Claude Max: advertised vs weekly multiple of Pro

Multiple of Pro's allowance, Anthropic's own figures, Aug 2025 – Jan 2026

Claude Max: advertised vs weekly multiple of Pro
LabelValue
Max 5x, advertised5×
Max 5x, per week3.5×
Max 20x, advertised20×
Max 20x, per week6×
Weekly multiples are midpoints of Anthropic's estimates of Sonnet 4 hours a week: Pro 40–80, Max 5x 140–280, Max 20x 240–480 (both ends give the same multiples). Per dollar, weekly hours fall up the ladder: 3.0 on Pro, 2.1 on Max 5x, 1.8 on Max 20x.

Source: Anthropic via TechCrunch

Anthropic deleted all six per-plan numbers from its Claude Code plans article between mid-January and early February 2026. Today it says only that Pro and Max "have a five-hour session limit and a weekly limit". The new Max, at $100, is "5x Pro usage" every five hours and every week, so for the first time the multiple is stated for the week too. We looked at what each plan is worth at API prices, and at outside measurements of the weekly gap, in Claude Code and Codex: subscription or API?.

Every big plan now has a week#

In 2024 most plans had one limit: messages per few hours, or requests per month. By November 2026 almost every vendor had stacked a second, longer cap on top, and several had switched the unit from messages to money.

From message counts to windows and credits

  1. Mar 2025

    Windsurf replaces "unlimited" with credits

    Codeium's $12 plan becomes Windsurf Pro at $15, with credits.
  2. 16 Jun 2025

    Cursor Pro: from 500 requests to $20 of usage

    Model usage billed at API prices; Cursor estimated about 225 Sonnet 4 requests.
  3. 18 Jun 2025

    Copilot enforces premium requests

    300 a month on Pro, 1,500 on Pro+, $0.04 each after that.
  4. 28 Aug 2025

    Claude adds weekly limits

    On top of the five-hour window, for Pro and Max.
  5. Sep 2025

    Codex: five-hour window plus a weekly limit

    Plus 30–150 local messages per 5 h, Pro 300–1,500.
  6. Nov 2025

    Codex sells credits past the limit

    They "kick in once you hit the 5-hour or the weekly usage limit".
  7. Mar 2026

    Windsurf moves to daily and weekly quotas

    Now Devin Desktop; credits stay only on legacy Enterprise plans.
  8. Apr 2026

    Codex credits priced per token

    From a per-message rate card to API-style token prices.
  9. 1 Jun 2026

    Copilot moves to AI credits

    One credit is $0.01 of tokens at API rates.
  10. Jun 2026

    Gemini app goes compute-based

    Refreshes every 5 hours until a weekly limit.
  11. Oct 2026

    Cursor's X Subscription: a weekly pool

    Claude, GPT and Gemini at API list price, on Premium and up.
  12. Oct 2026

    Claude Ultra: no five-hour limit

    10x Pro a week; Free gets a weekly limit too. First archived 1 Nov.

Source: host0 changelog of archived vendor pages; dates are announcement dates or the first capture that shows the change

Two models came out of this. Anthropic, OpenAI, Google and Windsurf sell time windows: a short rolling allowance plus a weekly cap, with the size of each stated loosely or as a multiple. GitHub and Cursor sell metered usage: an allowance in dollars of tokens at API prices, which is precise about money and vague about how much work it buys. Several vendors now combine both. Claude sells extra usage past the limit and gives Max and Ultra monthly API credits; Claude Enterprise is $20 a seat plus "usage at API rates". We covered the wider move from unlimited plans to credits in AI pricing in 2026.

Multipliers do the work#

When the unit is a request, the real limit is set by how much each model costs per request, and vendors change that without touching the headline number. Cursor's Pro plan went from 500 premium requests for $20, or 25 a dollar, to $20 of API-priced usage that Cursor said, "based on median token usage", covered about 225 Sonnet 4 requests: about 11 a dollar.

GitHub's docs history shows the same lever at its clearest. From 1 June 2026, most Copilot subscribers moved to AI credits. Annual Pro and Pro+ subscribers who stayed on request billing got a new table of model multipliers instead. Of 16 models in both tables, 12 went up and none went down; the median rose 4.3 times.

Copilot model multipliers for annual subscribers

Premium requests charged per interaction, before → from 1 Jun 2026 (annual Pro and Pro+ on request billing)

Copilot model multipliers for annual subscribers
LabelValue
GPT-5.4 mini0.33 → 6
Claude Sonnet 4.61 → 9
Claude Opus 4.63 → 27
Claude Sonnet 4.51 → 6
GPT-5.41 → 6
Gemini 3.1 Pro1 → 6
Claude Opus 4.53 → 15
Claude Opus 4.77.5 → 27
GPT-5.21 → 3
Bar length is the rise (after ÷ before). The same 300 requests a month buy a ninth as many Claude Sonnet 4.6 interactions. Four models were unchanged, including Claude Haiku 4.5 and Gemini 2.5 Pro. Monthly subscribers moved to AI credits, which have no published conversion to requests.

Source: host0 analysis of github/docs git history

Cursor's forum shows how the complaints followed the unit. Until mid-2025, titles were about request counts: "500 fast requests", the "slow pool". After the switch to dollars of usage they were about billing. Titles about plan limits peaked at 36.6 per 1,000 new topics in January 2025 and fell to 2.6 in November 2025. Pricing and billing titles peaked at 37.9 per 1,000 in July 2025, the month Cursor wrote that its changes "were not communicated clearly, and we take full responsibility" and offered "usage refunds for those with unexpected usage between June 16 and July 4".

Cursor forum: what people complain about

New topics per 1,000 whose title is about plan limits or about pricing and billing, Jan 2024 – Oct 2026

Cursor forum: what people complain about
Seriesxy
Plan limitsJan 20240
Plan limitsFeb 202419
Plan limitsMar 202431.2
Plan limitsApr 202416.3
Plan limitsMay 202432.9
Plan limitsJun 202420.8
Plan limitsJul 202417
Plan limitsAug 202411
Plan limitsSep 202411.2
Plan limitsOct 202411.7
Plan limitsNov 202425.3
Plan limitsDec 202420.7
Plan limitsJan 202536.6
Plan limitsFeb 202516.3
Plan limitsMar 202513.6
Plan limitsApr 202522.2
Plan limitsMay 202526.2
Plan limitsJun 202525.2
Plan limitsJul 202522.5
Plan limitsAug 202510.4
Plan limitsSep 20256.9
Plan limitsOct 202510.1
Plan limitsNov 20252.6
Plan limitsDec 20259.5
Plan limitsJan 202610.6
Plan limitsFeb 202611
Plan limitsMar 20263.8
Plan limitsApr 20267.1
Plan limitsMay 20268.1
Plan limitsJun 20269.8
Plan limitsJul 202613.5
Plan limitsAug 202611.7
Plan limitsSep 202618.8
Plan limitsOct 202613.6
Pricing and billingJan 20246.4
Pricing and billingFeb 202412.7
Pricing and billingMar 202415.6
Pricing and billingApr 202421.7
Pricing and billingMay 20249.4
Pricing and billingJun 202426
Pricing and billingJul 202417
Pricing and billingAug 202422
Pricing and billingSep 202419.3
Pricing and billingOct 202419.9
Pricing and billingNov 20248.4
Pricing and billingDec 202418
Pricing and billingJan 202513.5
Pricing and billingFeb 202516.8
Pricing and billingMar 20259.9
Pricing and billingApr 202510.1
Pricing and billingMay 202522.1
Pricing and billingJun 202529.4
Pricing and billingJul 202537.9
Pricing and billingAug 202528.7
Pricing and billingSep 202516.5
Pricing and billingOct 202517.1
Pricing and billingNov 202514.1
Pricing and billingDec 202514.6
Pricing and billingJan 202617.6
Pricing and billingFeb 202617.5
Pricing and billingMar 202615.8
Pricing and billingApr 202616.8
Pricing and billingMay 202616.9
Pricing and billingJun 202614.4
Pricing and billingJul 202624.5
Pricing and billingAug 202619.1
Pricing and billingSep 202618.2
Pricing and billingOct 202621.6
Plan-limit titles peaked at 36.6 per 1,000 in January 2025, in the request-count era; pricing and billing titles peaked at 37.9 in July 2025, after the switch to dollars. Moderators merge complaints into long threads, so title counts understate spikes, and 2024's small volumes are noisy.

Source: host0 analysis of forum.cursor.com

The top of the ladder went up#

As limits got vaguer, the plans got bigger. In January 2024 the most one person could pay any of these six vendors was $20 a month. By October 2026 Anthropic, Cursor and Windsurf sold a $200 plan, Google one at $199.99, and OpenAI one for $500.

Highest monthly plan one person can buy

US dollars a month, as of October 2026, with the earliest archived price in the note

Highest monthly plan one person can buy
LabelValue
OpenAI$500 (was $20)
Anthropic$200 (was $20)
Cursor$200 (was $20)
Windsurf$200 (was $12)
Google$199.99 (was $19.99)
GitHub$100 (was $10)
ChatGPT Pro went to $200 in December 2024 and $500 in September 2026. Google's AI Ultra launched at $249.99 in May 2025 and was cut to $199.99 a year later. Google's earliest is Gemini Advanced (February 2024), Windsurf's is Codeium Pro (August 2024); the others are January 2024.

Source: host0 analysis of vendor pricing pages (top_tier_prices.csv)

The top plans are where the limits are least specified. Claude Ultra is "10x Pro usage every week" with no five-hour limit for verified accounts. OpenAI's Pro tiers have no five-hour limit, only weekly ones. None of the $200 plans we read states its allowance as a number.

Complaints moved to the week#

Anthropic said its weekly limits would affect "less than 5% of subscribers based on current usage", and gave its reason: "Some of the biggest Claude Code fans are running it continuously in the background, 24/7," and "one user consumed tens of thousands in model usage on a $200 plan." The Claude Code repository shows how many people ran into the week anyway.

Claude Code issues with "weekly limit" in the title

Issues opened per month in anthropics/claude-code, title match only, Feb 2025 – Oct 2026

Claude Code issues with "weekly limit" in the title
xy
Feb 20250
Mar 20250
Apr 20250
May 20250
Jun 20250
Jul 20251 (caps announced)
Aug 20251
Sep 20250
Oct 20259
Nov 202512
Dec 20255
Jan 20269
Feb 202611
Mar 202619
Apr 202637
May 202630
Jun 202625
Jul 202631
Aug 202624
Sep 202631
Oct 202641
286 issues in all. The phrase never appears in a title before the weekly caps were announced in July 2025. Counting issues that use it anywhere in the title or body gives 1,014, an upper bound only: that looser match is about 70% precise.

Source: host0 count via the GitHub search API

Issue volume in both agent repositories exploded over the period (Claude Code went from 232 issues a month to more than 13,000 at its peak), so for the trend we divide by it. Titles with "rate limit", "usage limit" or "weekly limit" were 1,582 of Claude Code's 103,224 issues since February 2025. They peaked in March 2026, the month Anthropic tightened peak-hour limits: 281 issues, 24.8 per 1,000.

Limit complaints per 1,000 issues

Issues whose title has "rate limit", "usage limit" or "weekly limit", per 1,000 issues opened, by month

Limit complaints per 1,000 issues
Seriesxy
Claude CodeFeb 20250
Claude CodeMar 20250
Claude CodeApr 20250
Claude CodeMay 20255.6
Claude CodeJun 202511.7
Claude CodeJul 202518.2
Claude CodeAug 20258.2
Claude CodeSep 20258.2
Claude CodeOct 202518.5
Claude CodeNov 202515.3
Claude CodeDec 20258.4
Claude CodeJan 202612.2
Claude CodeFeb 202617.5
Claude CodeMar 202624.8
Claude CodeApr 202617.1
Claude CodeMay 202617.5
Claude CodeJun 202617.8
Claude CodeJul 202613.7
Claude CodeAug 20268.8
Claude CodeSep 202612.4
Claude CodeOct 202612.9
CodexAug 202524.4
CodexSep 202524
CodexOct 202520.3
CodexNov 202518
CodexDec 202513.4
CodexJan 202622.7
CodexFeb 20269.8
CodexMar 202614.6
CodexApr 202621.7
CodexMay 202615.5
CodexJun 202619.6
CodexJul 202618.6
CodexAug 202611
CodexSep 202616.8
CodexOct 20267.2
Claude Code peaked at 24.8 per 1,000 in March 2026 and ended at 12.9 in October; Codex peaked at 24.4 in August 2025 and ended at 7.2. Codex starts in August 2025: before ChatGPT sign-in its "rate limit" issues were mostly API errors, and monthly volumes were tiny.

Source: host0 count via the GitHub search API

Codex's worst month since its launch by the looser title-or-body count was September 2026, when 475 of 6,070 issues (7.8%) mentioned a limit phrase. That month OpenAI paused new $200 Pro sign-ups, with Thibault Sottiaux writing that these plans "put the most strain on our systems", and at the end of the month halved the $200 plan's limits.

Cuts get the attention#

On Hacker News, stories about the limits and pricing of AI coding tools went from a rarity to a regular feature. We matched story titles that name one of the products and a limit or pricing term.

Hacker News stories about AI coding-tool limits

Stories per year whose title names a product and a limit or pricing term, 2024 – Oct 2026

Hacker News stories about AI coding-tool limits
LabelValue
202415 (0.7 per 1,000)
202596 (3.2 per 1,000)
2026327 (8.1 per 1,000)
The notes divide by all stories naming one of the products, so the rise isn't just more AI coverage. 2026 runs to October. About 33 in 40 sampled matches were about plan limits or plan billing.

Source: host0 analysis of Algolia HN Search

The two biggest threads were GitHub Copilot's move to usage-based billing (27 April 2026: 767 points, 553 comments) and "Claude Code weekly rate limits" (28 July 2025: 609 points, 705 comments). Raises barely registered.

Hacker News reaction to limit changes

Points across all limit stories in the week after each change (from the day before to six days after)

Hacker News reaction to limit changes
LabelValue
Copilot usage-based billing814 (Apr 2026)
Claude weekly limits754 (Jul 2025)
Claude 5-hour limits doubled553 (May 2026)
Claude peak-hour cut519 (Mar 2026)
Codex 5-hour window back404 (Aug 2026)
Codex token rate card335 (Apr 2026)
Copilot credits go live53 (Jun 2026)
Holiday 2x ends31 (Jan 2026)
Claude holiday 2x12 (Dec 2025)
Highlighted bars are raises, or the end of one. The exception, the May 2026 doubling, shared its headline with a compute deal with SpaceX. Copilot's announcement drew 814 points; the week the credits actually went live drew 53.

Source: host0 analysis of Algolia HN Search

The reasons vendors gave fall into a few groups: demand and capacity (Anthropic in July 2025: Claude Code "has experienced unprecedented demand since launch"), heavy or abusive use (running it around the clock, sharing accounts and reselling access), and cost per request growing with agentic work. Cursor: "New models can spend more tokens per request on longer-horizon tasks." GitHub: "a quick chat question and a multi-hour autonomous coding session can cost the user the same amount." None of the statements we found mentions margins. GitHub's "sustainable, reliable Copilot business" comes closest.

People started metering themselves#

If the vendor won't say how much you get, you can measure how much you use. ccusage, an open-source tool that reads Claude Code's local logs and reports your usage, including per five-hour window, went from 1,406 npm downloads in May 2025 to 880,682 in October 2026.

Monthly npm downloads of ccusage

Downloads per month of the ccusage package, May 2025 – Oct 2026

Monthly npm downloads of ccusage
xy
May 20251K
Jun 202568K
Jul 2025105K
Aug 2025178K
Sep 2025122K
Oct 202575K
Nov 202574K
Dec 2025106K
Jan 2026144K
Feb 2026114K
Mar 2026201K
Apr 2026187K
May 2026330K
Jun 2026420K
Jul 2026461K
Aug 2026384K
Sep 2026481K
Oct 2026881K
Against Claude Code's own downloads, ccusage went from 0.09% to 2.2%. October's jump of 83% has no explanation in this data. Downloads include CI and reinstalls, so read them as interest, not users.

Source: npm downloads API

What this means if you build with AI#

  1. Measure your own usage before you pick a plan. A week of your real workload, read from a local meter like ccusage or from the vendor's usage dashboard, tells you more than any plan page. Note the tokens per day and how close you came to each window.
  2. Compare plans by the week, not the session. The weekly cap is what stops a heavy user. Max "20x" was 6× Pro a week by Anthropic's own numbers. When a plan quotes a multiple, ask "per what window?"
  3. Prefer a published number, or the API, when cost has to be predictable. ChatGPT Plus still states 160 messages every three hours, Copilot states its allowance in dollars of tokens, and API billing is exact per token. A budget you can put in a spreadsheet beats a "5x" you can't check.
  4. Watch the multipliers and units, not just the price. Copilot's annual subscribers kept their price and request count while Claude Sonnet 4.6 went from 1 request to 9. Cursor's $20 went from 25 requests a dollar to about 11. Re-check when a new model launches or the unit changes.
  5. Save the page when you subscribe. Limits changed 31 times across these six vendors in the first ten months of 2026. A dated copy (the Wayback Machine's "save page now" works) is the only way to know what you were promised.
  6. Keep a fallback. Cuts arrive with little notice and land on the busiest days. Keep a second agent, or an API key with a spending cap, configured, so a hit limit is a switch rather than a stop.
  7. Don't pay a year up front for a limit that isn't written down. Annual plans lock the price, not the allowance.

Whichever plan or agent wrote it, host0 puts what it made at a private link you can share, with a shared database it doesn't need to set up.

Methodology#

We used public data only, collected up to 1 November 2026, and re-opened the main pages we quote before writing.

  • Pages. 19 help-center, pricing and docs pages from Anthropic, OpenAI, Cursor, GitHub, Google and Windsurf, each under all its successive URLs. For each we fetched the first Wayback Machine capture of every month from January 2024, reduced each capture to the lines that state a limit, and bisected any month-to-month difference down to two adjacent captures. Arquivo.pt filled gaps. GitHub Copilot and Gemini CLI were dated from the git history of github/docs and google-gemini/gemini-cli, which gives exact commit dates.
  • Dates. "First seen" is the first capture with the new text, so it's an upper bound on when a page changed; captures are 1 to 30 days apart. Announcement dates are kept separately, and changes are counted by announcement date where one exists. Six 2026 changes reuse sources from our earlier posts, some of them news reports.
  • Classification. Each change was coded for direction (tightened, loosened, mixed, neutral) by a rule written before counting, and for the precision of the new limit (exact, estimate, relative, none). Every quote in the changelog was checked word for word against the archived text.
  • Checkpoints. For nine plans we coded how the vendor's page stated the main limit in July 2025 and November 2026. "A number" means exact or estimate. The result depends on the plan list: dropping Codex and Gemini CLI gives 6 of 7 → 2 of 7.
  • Limits of the archive. The Wayback Machine holds few 2024 captures of help-center articles, so 2024 is under-counted. Pages rendered with JavaScript (parts of the Codex pricing page, the Gemini API rate-limit table, chatgpt.com/pricing) hide numbers from the archive. Vendors also change limits without changing a page, through promotions and peak-hour adjustments; those only appear through announcements.
  • GitHub issues. Counts from the GitHub search API, one query per repository and month, all with complete results. In a hand check, 41 of 41 sampled title matches were about plan or rate limits; title-or-body matches were right 74 times in 105, so we use them only as an upper bound.
  • Hacker News. Every story naming one of ten products (91,155), from Algolia's search API, split into weekly windows wherever a month hit its 1,000-result cap. A limit story's title names a product and a limit or pricing term (438 stories). Points and comments are as of 1 November 2026.
  • Cursor forum. All 42,806 public topics from forum.cursor.com, titles only; the plan-limit pattern was right on 37 of 40 sampled titles. npm: monthly sums from the npm downloads API.

The data pack, with every change, source, tier and quote, is in our repository at docs/research/2026-11-01-ai-plan-limits.md. The changelog (changes.csv), the series behind every chart and the scripts that compute them are in docs/research/2026-11-01-ai-plan-limits-data/.

Open questions#

  • How big is a "weekly limit"? No vendor in this set publishes its weekly cap in a unit you can check. Outside measurements exist, but they disagree and go out of date with each change.
  • How many subscribers hit the limits? Anthropic's "less than 5%" is the only figure we found, from July 2025. Nobody has published one since.
  • What did 2024 look like? The archive is thin before 2025, and the o-series caps OpenAI set in 2024 and 2025 aren't in our changelog.
  • Will the week stick? OpenAI dropped the five-hour window for Codex in July 2026 and brought it back for Plus in August; Claude Ultra dropped it in November. The next round may be weeks only, or a return to plain credits.
ResearchAI codingPricing