Both products cap usage on a rolling 7-day window. The window starts at the first use after the previous window expired, usage inside it is capped at one weekly cap, and once capped you are locked out until the window ends. Both companies hand out goodwill resets for outages, quality incidents, launches and milestones. They implement the reset differently, and the difference changes what it costs.
How a Claude-style and a Codex-style weekly limit reset differ| Reset style | Usage at reset | Window end | Effect on a capped-out user | Source |
|---|
| Claude | used → 0 | unchanged | A second full cap, burned in whatever days remain of the old window; then the scheduled new window starts. | Anthropic help 17007452, “What is a limit reset?”: “Your weekly limits still reset on their usual day and time.” Matches user reports: implicator.ai (Jul 16 2026) “fixed weekly schedules gave some subscribers hours to use the refill and others days”; Reddit: Friday-reset account recovered ~86% of a week, Thursday-reset account almost none. |
| Codex | used → 0 | now + 7 days | The lockout ends early; the next full cap starts now. Every user’s window is now synced to the reset moment. | OpenAI help 20001498: “refreshes your 5-hour and weekly Codex usage windows and changes your weekly reset date”; OpenAI help 20001507: “Your new weekly usage period starts with your first request... Your next automatic weekly reset is scheduled 7 days after that first request.” OpenAI_Support forum reply Sep 6 2026: “a full banked reset changes your weekly reset date.” |
For a steady user who spends 1/7 of the cap per day and never hits it, both resets do nothing. All of the cost comes from users who hit the cap.
A power user, worked through
Take a power user who burns a full cap within a day or two of each window opening. With no reset, their caps start on day 0, 7 and 14. A Claude-style reset on day 4 clears usage but leaves the window end at day 7: they burn a second cap in the days left, and the next window still opens on day 7. Caps land on day 0, 4, 7 and 14, one extra cap, kept for good.
A Codex-style reset on day 4 clears usage and restarts the window, so caps land on day 0, 4, 11 and 18. Count caps at any moment and the gain over no reset oscillates between one cap (day 4 to 7, 11 to 14, …) and zero (day 7 to 11, 14 to 18, …). Over the long run it averages to the fraction of a window the reset let the user skip, here 3 of 7 days. That is why a Codex reset is worth roughly half a Claude reset per capped-out user, and why the raw hourly series (Advanced → show raw hourly series) climbs, falls back and climbs again after each OpenAI reset. The chart’s main lines are the trailing 7-day mean, which averages over exactly one 7-day cycle, so the sawtooth cancels and the lines only rise or round off.
Where in the window the reset lands
Gain for a power user from one reset, in caps, by hours remaining in their window when it lands:
Power-user gain from one reset by hours left in the window, in caps| Hours remaining | <24 | <48 | <72 | <96 | <120 | <144 | ≤168 |
|---|
| Claude | 0.34 | 0.67 | 0.85 | 0.91 | 0.84 | 0.68 | 0.29 |
| Codex | 0.07 | 0.23 | 0.35 | 0.47 | 0.51 | 0.48 | 0.24 |
Claude’s gain ≈ “what you had already burned, capped by what you can re-burn before the window ends”. Codex’s gain ≈ “the lockout days you get to skip”.
Dates come from @ClaudeDevs and @thsottiaux posts, timestamped in UTC from the tweet snowflake IDs and cross-checked against claude-resets.com (14 events) and codex-reset.com /api/timeline (41 events since Sep 2025), plus 9 more @thsottiaux reset tweets the tracker JSON omits (Dec 25, Apr 7, Apr 9, Apr 17, Apr 28, Jul 12, Jul 29, Aug 12, Aug 27), 50 total. Two announcements less than 8 hours apart count as one reset. Anthropic did no all-user reset between the weekly limit’s launch (Aug 28 2025) and Apr 16 2026; OpenAI’s first tracked reset is Sep 17 2025.
Inside the simulated window, Apr 16 – Sep 22 2026: Anthropic 13 hard resets + 1 banked; OpenAI 34 hard + 5 banked. Targeted grants are not modeled: Jul 12 (500k users), Sep 3 (one per day without Astra access), and the Sep 9 re-issue of failed redemptions. These tables are rendered from the same event list the simulation runs on.
Anthropic
Every Anthropic (Claude) usage reset, Apr 16 – Sep 22 2026| Date (UTC) | Scope | Reason | Source |
|---|
Cadence: 2.7 per month, longest gap 47 days (Jul 16 → Sep 1).
OpenAI
Every OpenAI (Codex) reset since Sep 2025; rows outside the simulated window are dimmed| Date (UTC) | Type | Reason | Source |
|---|
Cadence: median 4.2 days between resets all-time, 4.9 days over the last 30. Eight resets in the 13 days Jul 9-21. One reset was a mistake (Sep 9: banked resets failed to apply, re-issued with an apology email) and one hard reset wiped users’ stacked banked resets (Jun 28). Banked rows link to the account because the grants were announced across several posts.
Model launches on the chart
The dotted lines on the chart, from each company’s own announcement, dated from the announcing tweet ID where one exists.
Model launches marked on the chart| Date (UTC) | Company | Launch | Source |
|---|
A Monte Carlo model of a subscriber base, stepped hour by hour. Every user has a weekly demand and a daily pattern; the weekly cap and each reset mechanic are applied exactly as documented.
Simulated population by segment| Segment | Share | Weekly demand (in caps) | Shape |
|---|
| Casual | 55% | lognormal, median 0.12 | random active days |
| Regular | 27% | lognormal, median 0.45 | weekday-heavy, 1-2 spike days at 3x |
| Heavy | 15% | lognormal, median 1.1 | weekday-heavy, 1-2 spike days at 3x |
| Power / account jugglers | 3% (swept 1-5%) | Pareto, xmin 2, α 1.3, capped at 12 | every day, 14-18h active; burns a week’s cap in under 2 days and would burn more if allowed |
Result: 13% of users hit the weekly cap in a given week (0.2% casual, 6% regular, 57% heavy, 100% power). Anthropic said in Jul 2025 that weekly limits would “apply to less than 5% of subscribers”; the model is deliberately heavier than that, so the dollar figures are an upper-middle estimate. Average consumption is 0.37 caps per user per week.
The clock
- Rolling 7-day window from first use; usage inside it is capped at one weekly cap; a capped user is locked out until the window ends.
- Hourly steps, 3 weeks of burn-in before the measured window, a random first-use hour per user.
- Timezones: each user is placed 50% Americas, 25% Europe, 25% Asia, with day-of-week and spike-day demand and gamma noise on daily and hourly demand.
- Common random numbers: the same users, with the same demand, run twice, with and without the resets. The gap is the giveaway, so the extra consumption is exact rather than a difference of two noisy runs.
- Banked resets are redeemed the first hour their holder is capped out, the most expensive way to use one.
Not modeled
- The 5-hour limit, model-specific caps, and extra-usage purchases (which would reduce the marginal cost of a reset for both).
- Plan mix within a segment, and account sharing: jugglers are modeled per account, their other accounts are other users in the population.
- Demand response: a reset does not make anyone use more than they would have if unlimited.
- The first-request anchor: a real Codex window restarts at the first request after the reset + 7 days, while the model uses reset time + 7 days. For cap-hitters, who are active within the hour, these coincide; idle users never hit the cap, so the cost impact is negligible.
This page runs the same model in a Web Worker in your browser. Checked against the Python reference at N = 20,000, seed 1, every scenario metric lands within 4.1% and the headline totals within 2.5%.
One reset, midweek (Wed 14:00, 3% power users)
Cost is extra consumption vs. no reset, in weekly caps, after the Codex-style window shift has settled (one-week period average).
One reset, midweek: Claude-style vs Codex-style | Claude-style | Codex-style | Ratio |
|---|
| Extra consumption, % of one week’s population usage | 12.4% | 7.4% | 1.7x |
| Extra caps per 1,000 users | 45 | 27 | 1.7x |
| Per power user | 0.66 cap | 0.34 cap | 2.0x |
| Per heavy user | 0.16 cap | 0.10 cap | 1.6x |
| Per regular user | 0.009 cap | 0.009 cap | 1x |
| Per casual user | 0 | 0 | |
| Share of the cost going to the top 3% of users | 57% | 62% | |
Stable across seeds (±0.2 points). Power-user share 1% → 5% moves the totals 9.5% → 15.0% (Claude) and 6.0% → 8.8% (Codex); the ratio stays 1.6-1.7x.
The real timelines, Apr 16 – Sep 23 2026, N = 20,000 subscribers
Cost is extra consumption vs. no resets, final-week average.
Real reset timelines, Anthropic vs OpenAI, N = 20,000| Scenario | Resets | Extra caps per 1,000 subs | Weeks of population usage | Per reset (% of a week) | Per power user | Per heavy user |
|---|
| Anthropic dates × Claude mechanic (actual) | 13 | 495 | 1.36 | 10.4% | 7.8 caps | 1.6 caps |
| Anthropic dates × Codex mechanic | 13 | 234 | 0.64 | 4.9% | 3.6 | 0.8 |
| OpenAI dates × Codex mechanic (actual) | 34 | 708 | 1.94 | 5.7% | 12.9 | 2.0 |
| OpenAI dates × Claude mechanic | 34 | 1,042 | 2.85 | 8.4% | 18.4 | 3.0 |
| OpenAI actual + 5 banked grants (“Tibo button”) | 34 + 5 banked | 921 | 2.52 | 11.6% per banked | 14.9 | 2.9 |
Over five months, an Anthropic power user got 7.8 extra weeks of quota out of 22 weeks; an OpenAI power user got 12.9. OpenAI’s cheaper mechanic was more than eaten by doing 2.6x the resets. Had OpenAI used Claude’s mechanic at its own cadence it would have given away 2.85 weeks; had Anthropic used Codex’s mechanic it would have given away 0.64.
Banked resets: the “Tibo button”
The 34 OpenAI events are hard resets only. On top of them OpenAI deposited banked resets into every paid account, redeemable whenever the user wants (community name: the “Tibo button”, after Tibo’s Aug 20 “gifted a very fancy new reset button” post). All-paid-user grants in the window: Jun 18 (“sneaky double reset”), Jun 29 (+1 banked after the Jun 28 hard reset wiped stacks of up to three), Jul 13 (7M milestone), Aug 21-22 (20M milestone), Sep 22 (GPT-6). Adding the five takes OpenAI from 1.94 to 2.52 weeks of usage, +0.116 weeks per banked grant: a banked Codex reset (11.6% of a week) is worth about two hard Codex resets (5.7%) and roughly one Claude reset (10.4%), because the user picks the moment. Anthropic’s single banked reset (Sep 22, Opus 5.5, redeem by Oct 22) lands after the measurement window, so it barely registers.
Who gets it
~60% of the cost goes to the top 3% of users under both policies (57% Claude-style, 62% Codex-style for a single reset). A reset is a gift to exactly the people the weekly limit was introduced to contain.
Timing
Claude-style barely cares when you reset (10-12%). Codex-style drops to 4.4% on a weekend vs 7.4% midweek, about 40% cheaper, because the synced window starts while heavy users are idle.
Synchronization
After each Codex-style reset, every cap-hitter’s window starts at the same moment, so their burn-then-lockout pattern lines up: a demand wave every 7 days. That is the oscillation in the raw hourly series (Advanced) and the lockstep in the OpenAI dot grid. Claude-style leaves the phase distribution alone, and its grid stays staggered.
Nobody publishes the cap in tokens or dollars, so the model uses measured API-equivalent ceilings:
Weekly cap values used for the dollar figures| Plan | Weekly cap, API-list value | Source |
|---|
| Claude Max 20x ($200) | ~$1,850/wk (low) to ~$5,200/wk (high) | SemiAnalysis Aug 23 2026: “$200/mo Claude plan… up to $8,000/mo” (tested Jun 2026); ksred.com cctrack log on Max 5x: $5,623/mo → scaled 4x for high |
| Claude Max 5x ($100) | ~$460 to ~$1,300/wk | 1/4 of 20x (low); ksred.com $5,623/mo (high) |
| Claude Pro ($20) | ~$90 to ~$260/wk | 1/5 of Max 5x |
| Codex Pro 20x ($200) | ~$3,230/wk | SemiAnalysis: “OpenAI’s $200/month plan could yield up to $14,000/month” |
| Codex Pro 5x / Plus | ~$750 / ~$160/wk | scaled by published message allowances (700 / 150 vs 3,000 per 5h) |
| OpenAI’s own price for a reset | Plus $5-8, Pro Lite $25-40, Pro $50-80 | purchasable “instant reset” (TestingCatalog Aug 7 2026; explainx “$80 usage reset on $200 Pro plan”); an instant reset “restores the applicable usage allowances immediately and starts a new weekly period” (help 20001507) |
Plan mapping: power users on the 20x plan, heavy on 5x, regular and casual on the base plan. Cost per subscriber = Σ segment share × extra caps per user × cap value.
Cost per subscriber, at API-list value | Per subscriber, Apr–Sep 2026 | Per subscriber per reset | Per 1M subscribers per reset |
|---|
| Anthropic, actual (13 resets) | $540 – $1,520 | $42 – $117 | $42M – $117M |
| Anthropic, had it used Codex mechanic | $250 – $710 | $20 – $55 | $20M – $55M |
| OpenAI, actual (34 resets) | ~$1,510 | ~$45 | ~$45M |
| OpenAI, had it used Claude mechanic | ~$2,190 | ~$64 | ~$64M |
At OpenAI’s own retail price. At $80 on Pro and $8 on Plus instead of API value, a global Codex reset is worth roughly $10 per subscriber (3% on Pro 20x at $80, 15% on Pro 5x at ~$40, rest on Plus at $8), i.e. about $10M per million subscribers per reset: $340M for the 34 hard resets, ~$390M for all 39 counting the banked grants. That is what OpenAI itself would charge; the API-list figure is what the usage would have billed.
The whole dollar figure is basically: power-user share × 20x-plan cap value × ~0.6 caps per reset (Claude) or ~0.35 (Codex). Change either input and it scales linearly. Neither company discloses subscriber counts: Claude Code was at $2.5B run-rate in Feb 2026 with >50% enterprise; OpenAI reports 55M paying ChatGPT subscribers (Q1 2026) and 25M Codex “active users” (Aug 31 2026, window undefined, includes free). One million paid subscribers is a plausible order of magnitude for either, which is why the chart is per 1M.
From list price to real compute: two routes
Route A, usage ratio. The resets added 1.36 weeks (Anthropic) and 2.52 weeks (OpenAI, 34 hard + 5 banked; 1.94 without the banked grants) of subscriber consumption on top of a 22-week window, i.e. +6.2% and +11.5% of subscriber inference for the period. Multiply by what each company spends serving subscribers.
Route B, API-list value × list-to-cost ratio. The chart’s $544M / $1.51B per million subscribers, times the real cost of a list-price token. Anthropic told investors inference gross margin is “above 80% before revenue share and training” (FT, Sep 13 2026), so real cost is under 20% of list; Martin Alderson’s Mar 2026 estimate is ~10%. Use 10-20%.
Burn and inference anchors (all The Information / FT / Bloomberg / Reuters via syndication; primary pages paywalled):
Burn and inference anchors | OpenAI | Anthropic |
|---|
| 2025 inference spend | $8.4B (4x YoY; $3.9B of it on free users) | $2.7B (3x YoY) |
| 2026 inference | ~$14.1B projected (secondary source) inside ~$50B total compute | Jan plan $7B inference + $12B training on $18B revenue, now stale: Q2 revenue alone was $11.5B and compute ran $0.56 per revenue dollar (~$6.4B/quarter all-in) |
| Cash burn | $9B (2025), $25-27B forecast 2026; Q1 2026 burned $3.7B on $5.7B revenue | ~$3B (2025); adjusted-operating-profitable since Q2 2026 ($559M on $10.9B), so “burn” is the wrong denominator, inference spend is the right one |
| Gross margin | 33% (2025) → 39% (Q1 2026) | 40% (2025) → ~44% (2026); >80% on inference alone |
| Subscriber base | 55M paying ChatGPT (Q1 2026); Codex >5M WAU (Jun), “25M active” (Aug 31, window undefined); Pro <1% of subs | undisclosed; “more than doubled” in 2026, majority on Pro |
| Consumer share of revenue | <50% as of Aug 2026 (enterprise crossed over) | ~15-20% (80-85% enterprise API) |
Over the 22-week window that is roughly $5.9B of inference for OpenAI ($14.1B × 22/52) and $4-5.5B for Anthropic (Q2 pace, inference-only ≈ 20% of revenue on a $50-65B run-rate).
Estimated real compute cost of the resets, Apr 16 – Sep 22 2026| Real compute cost, Apr 16 – Sep 22 2026 | Anthropic (13 resets) | OpenAI (34 hard + 5 banked resets) |
|---|
| Route A: subscriber inference in window | $1-2.6B (subs ≈ 15-20% of ~$22B revenue in window, inference at 30-60% of sub revenue) | $1-1.9B (paid-user inference ≈ 54% of $5.9B, Codex + Work at 30-60% of that) |
| Route A: reset cost | $60-160M (6.2%) | $115-220M (11.5%; $90-170M for the 34 hard resets alone) |
| Route B: real cost at 10-20% of list | $55-110M per 1M subs; $110-550M at 2-5M Claude Code-capable subs | $150-300M per 1M subs; $450M-2.4B at 3-8M paid Codex users |
| As % of inference spend in window | ~1.5-10%, central ~3% | ~2-4% by route A; route B’s top end exceeds all paid-user inference and is not credible |
| As % of cash burn in window | n/a (profitable) | ~1-2% of $10.5-11B |
Why route B runs hot, and why route A is the number to quote. The simulated population (13% cap-hitters, 3% jugglers at $3,230 a week) implies ~$950/month of list-price usage per average subscriber, which no subscription business survives even at 20% list-to-cost. OpenAI’s own description of Plus users as “relatively casual” and Anthropic’s “<5% hit the weekly limit” both say the real base is lighter than the sim. Route A does not depend on that assumption: on the order of $100M of real compute for Anthropic and $100-200M for OpenAI over five months, low single-digit percent of inference spend, and for OpenAI 1-2% of burn. Cheap as a retention lever; expensive as a habit, since Tibo’s cadence (a reset every 4-5 days) compounds it and every reset is a gift to the top 3%.
What would move it: an actual Pro/Max or Codex-paid subscriber count (nobody has published one), or any “top N% of users = M% of inference” statistic (none exists in print).
Checked Sep 23 2026 (Wayback captures of the help center, openai/codex GitHub issues, OpenAI community forum, Tibo’s posts). Every primary source from Oct 2025 to Sep 2026 says the same thing: a Codex reset, whether global/hard, banked or paid, puts usage back to 100% and restarts the weekly window, anchored to the first request after the reset plus 7 days. What changed was delivery and documentation, not semantics.
Did OpenAI change the Codex reset mechanic? Dated primary sources| Date | Source | What it says |
|---|
| Oct 30-31 2025 | openai/codex#5999 | Weekly reset moved Nov 3 → Nov 7 after a goodwill reset; “the goalposts move by five days” |
| Nov 24 2025 | #5999, etraut-openai | “When a service outage occurs, we typically reset rate limit dates as a way of crediting users… We are considering your idea [don’t move the anchor]” (never shipped) |
| Mar 14 2026 | #14701, etraut-openai | “it resets the 5-hour and 7-day time windows for all users, and everyone’s ‘remaining usage’ goes back to 100%” |
| May 23 2026 | #9508 comment | Users contrast: “When Claude resets, they keep the weekly reset time unchanged” |
| Jun 12 2026 | @OpenAI, @thsottiaux | Banked resets launch: “you will get to choose when it actually applies” (changes when, not what) |
| Jun 28 2026 | @thsottiaux | “hard reset” = pushed to everyone, vs redeemable; same mechanic |
| Jul 10 2026 | community.openai.com/t/1364615, OpenAI_Support | “After a reset, the next seven-day window starts with the next Codex use, so the date shown can move” |
| Jul 29 → Aug 13 2026 | Wayback, help 11369540 | Sentence “your weekly reset date will move to roughly seven days after you redeem it” first appears Aug 13 |
| ~Aug 31 2026 | OpenAI help 20001507 | Paid instant reset: “next automatic weekly reset is 7 days after that first request” |
| Sep 4 2026 | OpenAI help 20001498 created | “Using a full banked reset refreshes your 5-hour and weekly Codex usage windows and changes your weekly reset date” |
| Sep 6-8 2026 | OpenAI_Support forum replies | “a full banked reset changes your weekly reset date”; a fixed reset day is “a scheduling suggestion; we don’t have a timeline” |
| Sep 8-10 2026 | #43643, #43694, #44538 | Sep 7 global reset “overwriting users’ current usage window… including resets announced by Tibo” |
Counterexamples are partial failures, not a different mechanic: #33344 (reset never applied, so nothing moved), #32421 and #34874 (timer moved, usage not restored), #44201 (one report of the anchor staying put after Sep 7). The Sep 9 2026 status incident (“unexpected usage limit resets”, ~25 min) produced dozens of date-jump reports that are noise, not policy. Business workspaces show repeated misses (#27027 Jun 4, #33344 Jul 14), so a small per-reset miss rate would lower the Codex figure slightly.
Modeling consequence: one mechanic for all 34 hard resets, no split.
Anthropic side. Anthropic help 17007452 (“Your weekly limits still reset on their usual day and time”) has no Wayback history and describes the user-redeemed reset of Sep 22 2026. Whether Anthropic’s 13 global resets in the window also kept the reset day is being verified separately; the model assumes they did.
Elvis Sun, founder at Medialyst and AI engineer. Follow along on X at @elvissun and on LinkedIn. The simulation, the reset dataset and this page are his; numbers are a simulation, not disclosed data.
The chart’s main line is the trailing 7-day mean of the extra usage, which only rises; the raw hourly series (turn it on under Advanced → show raw hourly series) is what dips. The raw series is cumulative extra usage compared with a no-reset world. A Codex-style reset restarts every user’s 7-day clock at the same moment, so everyone who was locked out burns again at once and the gap climbs. Two days later the heavy users cap out together and sit locked out while, in the no-reset world, their staggered windows were renewing. For that stretch the reset world uses less, so the gap shrinks. Seven days after the reset every synchronized window renews at the same hour and it climbs again. The sawtooth fades as random demand desynchronizes people. Part of what a Codex reset gives is borrowed from the following week and handed back in the trough; the 7-day mean nets that out.
A Claude reset clears usage but leaves each user's window end where it was. The reset world and the no-reset world stay in phase, so the extra usage is added and never given back. One step per reset, flat in between.
Because of the give-back above. At the moment of the reset both feel identical: a fresh cap. Under Claude your regular reset still lands on its usual day, so you get another cap on top. Under Codex your next reset moves to 7 days after the goodwill one, so you wait longer. Per reset in this window: ~10.4% of a week of population usage (Claude mechanic) vs ~5.7% (Codex).
Only people who were locked out. Roughly 60% of every reset's value, under both mechanics, goes to the top ~3% of users who burn a week of quota in under two days. A casual user who never hits the cap gets nothing from either.
From @ClaudeDevs and @thsottiaux posts, timestamped from the tweet IDs, cross-checked against claude-resets.com and codex-reset.com. Two announcements less than 8 hours apart count as one reset. Window: Apr 16 to Sep 22 2026, 13 Anthropic hard resets + 1 banked, 34 OpenAI hard resets + 5 banked.
No. A reset sets usage to zero; a second one hours later only helps users who burned quota in between, which is almost nobody. The per-reset figures are averages over the real cadence, back-to-back ones included.
Real: the reset dates, kinds, and both mechanics as documented in each company's help center (Anthropic 17007452, OpenAI 20001498 / 20001507). Simulated: the population (55% casual, 27% regular, 15% heavy, 3% power, Pareto tail), hourly demand with timezones and weekday patterns, and the rolling 7-day window with lockout at the cap. Not modeled: the 5-hour limit, plan mix within a segment, account sharing. The same random demand runs with and without the resets; the difference is the giveaway.
It is a Monte Carlo simulation. Different seeds give different random users; at N = 6,000 the headline figures wobble by a few percent, at N = 20,000 less. The ratios between the two companies barely move.
General-availability model launches, taken from each company's own announcement and dated from the announcing tweet ID where one exists (otherwise the announcement date). Anthropic launches are terracotta, OpenAI launches light gray. Resets often cluster right after a launch, when new-model demand pushes people into their limits.
The worked example under Two weekly limit resets, one difference is the whole story. A Codex-style reset hands a capped-out user a cap early, then pushes their next window back by the same amount, so their gain over the no-reset world swings between one cap and zero. Cumulative usage never goes down: during the trough the reset world simply stops growing while the no-reset world, whose windows renewed on schedule, catches up, so the net gap shrinks. Summed over everyone whose clock was synchronized, that is the sawtooth in the raw hourly series. The main line, the tip labels, the panel numbers and the tooltip all use the trailing 7-day (168-hour) mean, which averages over exactly one cycle, so the sawtooth cancels and the line only rises or rounds off. It starts on Apr 16 from the hours available, rather than waiting a week. Residual noise in the mean is held at its running high, so the plotted value never ticks down; the largest dip that hides is under 0.1% of the final value.
Because cadence and mechanic interact. A Codex reset is worth about half a Claude reset per capped-out user (5.7% vs 10.4% of a week in this window), so OpenAI’s 34 are not 2.6x Anthropic’s 13. And back-to-back resets are worth less than spaced ones: a reset only helps people who burned quota since the last one, so a second reset hours after the first gives almost nothing.
For a Codex-style reset, yes: a single reset on a weekend costs 4.4% of a week vs 7.4% midweek, about 40% cheaper, because the synchronized window starts while heavy users are idle. A Claude-style reset barely cares when it lands (10-12%).
A real paid-subscriber count for Claude Code or Codex, or any “top N% of users = M% of inference” statistic. Neither has been published. The dollar figures scale linearly with the power-user share and the 20x-plan cap value.
No, only the weekly cap. Modeling it would slightly reduce what a power user can extract from a Claude-style reset late in the window, since the 5-hour limit caps how fast a second weekly cap can be burned.
Yes. Every slider re-runs the same model in your browser, and the random seed and population size N are under Advanced in the left panel. This page’s port matches the Python reference within ~4% at N = 20,000.
The chart uses the low API-list anchors: Anthropic’s 13 resets = $544M and OpenAI’s 34 hard resets = $1.51B per million subscribers, with the counterfactuals at $258M and $2.19B. Adding the 5 banked grants takes OpenAI’s line to about $1.85B at default settings.
- The Codex reset is not free. Restarting the timer makes a reset worthless to steady users, but they cost nothing under either policy. For a capped-out user, a Codex reset ends the lockout early, worth about half of Claude’s double-cap. Ratio 1.7x overall, 2x on power users. Not 10x.
- Cadence beats mechanic. OpenAI’s cheaper reset, done 34 times plus 5 banked grants, cost far more than Anthropic’s expensive one done 13 times (2.5 vs 1.4 weeks of usage, 1.9 vs 1.4 on hard resets alone; ~$45 vs ~$42 per subscriber per reset at list, so ~$1,510 vs ~$540-1,520 per subscriber over five months).
- The cost is concentrated. ~60% of it goes to the top 3% of users under both policies. A reset is a gift to exactly the people the weekly limit was introduced to contain.
- Codex-style resets synchronize the population. After each one, every cap-hitter’s window starts at the same moment, so their burn-then-lockout pattern lines up: a demand wave every 7 days. Visible as the oscillation in the raw hourly series (Advanced). Claude-style leaves the phase distribution alone.
- If you must do a Claude-style reset, timing does not help. If you do a Codex-style one, do it on the weekend (~40% cheaper).
- Banked resets are the most expensive variant of either mechanic because the user chooses the moment: a banked Codex reset is worth ~2 hard Codex resets, about one Claude reset. OpenAI handed out five to every paid user (Jun 18 to Sep 22); Anthropic’s first was Sep 22 2026.
- Real compute cost is ~$100M per company over five months, low single digits of inference spend. The $500M-$1.5B chart numbers are API-list value, what the usage would have billed, not what it cost to serve.
Model caveats
- Segment shares and demand distributions are assumed. The model’s 13% weekly cap-hit rate is above Anthropic’s stated “<5%”, so scale dollar figures down if you believe them.
- Cap values changed during the window (Claude +50% promo May 13 – Sep 14, then +25% permanent; Codex 5-hour limits dropped Jul 12 and restored for Plus Aug 25). Treated as constant.
- Ignores the 5-hour cap, model-specific caps, and extra-usage purchases (which would reduce the marginal cost of a reset for both).
- Jugglers are modeled per account; their other accounts are other users in the population.
- No demand response: the reset does not make anyone use more than they would have if unlimited.
- Both mechanics are documented by the companies’ help centers (Anthropic 17007452, OpenAI 20001498 / 20001507). Two Codex counterexamples exist (#33344: a Business workspace’s reset date did not move on Jul 14; #32421: timer moved but usage stayed at 55%), so the Codex mechanic is not perfectly consistent.
Primary sources
- Anthropic help center 17007452, “What is a limit reset?”
- OpenAI help center 20001498 (banked resets), 20001507 (instant reset, weekly period), 11369540 (via Wayback)
- Reset announcements: @ClaudeDevs and @thsottiaux on X (every reset links to its post in Every Claude usage reset and Codex reset, dated)
- Trackers: claude-resets.com, codex-reset.com (
/api/timeline)
- openai/codex GitHub issues: #5999, #9508, #14701, #27027, #32421, #33344, #34874, #43643, #43694, #44201, #44538
- OpenAI community forum: thread 1364615 and OpenAI_Support replies of Sep 6-8 2026
Secondary sources
- secondary SemiAnalysis, Aug 23 2026: subscription value vs API list price (“$200/mo Claude plan… up to $8,000/mo”; “OpenAI’s $200/month plan could yield up to $14,000/month”)
- secondary ksred.com cctrack log: Claude Max 5x at $5,623/mo of API-equivalent usage
- secondary Financial Times, Sep 13 2026: Anthropic inference gross margin “above 80% before revenue share and training”
- secondary The Information, Bloomberg, Reuters (via syndication; primary pages paywalled): OpenAI inference spend, gross margin, cash burn; Anthropic revenue and compute
- secondary PitchBook, Jun 2026: Anthropic compute per revenue dollar ($0.56-0.71 in H1 2026)
- secondary Martin Alderson, Mar 2026: real cost ≈ 10% of list price
- secondary TestingCatalog, Aug 7 2026, and explainx: prices for OpenAI’s purchasable instant reset
- secondary implicator.ai, Jul 16 2026: users on fixed weekly schedules after a Claude reset
- secondary Aron Prins, post of Sep 22 2026 on X: the Claude vs Codex reset debate this page answers
- secondary “Tibo button” nickname: @AGTPinsights
Links without a URL here were not recorded with a stable public address; they are cited by outlet and date. Dollar figures of $1,000M or more are shown in billions throughout the page.
The Reset War · back to the chart · elvis.so/reset-war · made by @elvissun on X