August 20: purchased full resets appear on some accounts

Multiple dated community reports now show a purchased full reset offer on some Codex accounts. The observed prices are $8 on some Plus accounts and $80 on some Pro 20x accounts. These prices and the rollout remain community-observed: OpenAI has not published universal pricing, eligibility, plan coverage, refund terms, or availability for this control.

The reported mechanism matters more than the headline price. Redeeming the offer starts a new seven-day window and moves the next scheduled weekly reset to roughly seven days after redemption. It is therefore not a bonus week stacked on top of the old recovery date. Pro 5x, Business, universal availability, refund behavior, and exact eligibility remain unresolved.

This evidence does not show that paid resets replaced free promotional or incident resets. Provider-wide hard resets, banked promotional resets, purchased full resets, purchased credits, and scheduled recovery remain separate mechanisms. Check the signed-in usage page before making a purchase and compare the account’s displayed price with the value of waiting, switching lanes, or buying credits.

August 13: a 15M-milestone reset, not a new entitlement

At 01:01 UTC on August 13, 2026, OpenAI product lead Tibo Sottiaux said Codex had crossed 15 million active users and told “everyone” to enjoy a reset expected within roughly an hour. The direct announcement establishes the event and its milestone framing. It does not create a permanent reset schedule, a banked grant, or an every-million entitlement.

The distinction matters because the prior public promise ended at 10M. On August 12, Sottiaux said explicitly that he had promised a reset for every additional million active users only through 10M, then teased a surprise after growth passed that point.

The recent provider-wide sequence is unusually dense:

For future teases, the practical feed is Tibo Sottiaux on X. Treat it as a volatile announcement stream, not a policy changelog or a substitute for the signed-in usage page.

Non-rollover capacity has an expiry cost

An AI subscription meter is not a wallet. It grants service capacity under a provider’s current rules, often inside weekly, session, model-specific, or credit-based accounting. If unused allowance does not roll over, it is perishable: scheduled recovery or a provider-wide reset can erase the opportunity to use what remained.

Ration too aggressively and an expiry or surprise reset can erase useful option value. Spend too aggressively and you may lack capacity for urgent work. The goal is useful accepted work per subscription period, with a reserve—not the highest possible token burn.

Six mechanisms that should not be mixed

MechanismWho controls itWhat happens
Scheduled recoveryPublished plan clockThe old window ends; unused allowance normally does not roll over
Provider-wide hard resetProviderThe announcement’s group receives a fresh meter automatically
Incident or promotional resetProviderScope and reason depend on the named event
Banked resetEligible user until expiryThe user chooses when to refresh the applicable limit
Purchased full resetEligible purchaser; account-scoped rolloutCommunity reports describe a fresh seven-day window whose next recovery date moves with redemption
Purchased usage creditsAccount or workspace purchaserWork can continue after included limits; counters are not documented as refilled

The labels protect against false equivalence. A broad hard reset is not cash or a credit deposited into an account. A banked reset is expiring functionality, not permanent rollover. OpenAI pricing says available credits let work continue after included limits; it does not say purchased credits restart a five-hour window or refill a weekly meter.

Five-hour windows: documented, while observed accounts show weekly-only

OpenAI’s pricing page still describes shared five-hour windows and possible additional weekly limits. That is the documented policy surface.

Observed account behavior is different from that documentation. Current reports continue to show weekly-only displays after the announced July 30 restoration date. Their current enforcement and future policy state are unresolved/uneven because OpenAI still documents the windows, has not published a deprecation notice, and account behavior can differ. “Active for every account,” “not in use anywhere,” and “permanently removed” would each overstate the evidence.

AIHackers does not expect dependable five-hour enforcement to return in the near term. That is a dated forecast, not an OpenAI commitment. A product update or first-party announcement could falsify it immediately.

Banked resets: the broad July grant reached its expiry window

OpenAI’s June 11 release notes say banked resets are usable for 30 days after grant. On July 12 at 21:28 UTC—July 13 in Singapore—Sottiaux announced a 500,000-user grant and said the first broad banked reset for ChatGPT Work and Codex users would follow for the 7M milestone.

No later directly announced broad banked grant was found in the bounded August 13 source review. A grant made around July 13 would therefore reach its 30-day window around August 12–13, subject to the exact grant timestamp and applicable offer terms.

Dated Reddit reports describe grants disappearing near or before the displayed expiry and, in some cases, without the notification users expected. Those reports are community-observed, not proof of universal behavior, backend intent, or a formal end to banked resets. Current referral offers can still grant a reset, usage, or another benefit under offer-specific terms.

Game theory of frequent surprise resets

Frequent broad resets can function as an acquisition and retention subsidy. They lower the experienced cost of trying a demanding model, keep rate-limited users active, and can help form a habit before standard limits become the dominant constraint. Tying resets to user milestones also creates a compact X marketing loop: growth produces an announcement, the announcement produces attention, and the reset gives current users a reason to reopen the product.

If spare compute exists at particular times, a discretionary reset may improve reserve-capacity utilization. But the same event can raise aggregate burn by weakening the incentive to ration. It can also create adverse selection: power users are most likely to watch the feed, exhaust allowance before a teased reset, and capture a larger share of the subsidy.

That creates reset moral hazard. Once users expect another refresh, conserving capacity becomes less attractive; some community posts explicitly describe accelerating usage before a promised reset. A forced reset can also destroy remaining allowance for users who did conserve, which makes timing feel arbitrary and encourages even more anticipatory burn next time.

The likely pressure is toward a more explicit accounting system: transparent credits, user-controlled grants, fixed promotional boosts, or tighter and more predictable limits. None of those outcomes is guaranteed. The key falsifier is whether broad discretionary resets remain frequent without a move toward clearer metering or narrower eligibility.

Competitor promotions use different accounting

ProviderCurrent promotion or changeWhat it is not
AnthropicPro, Max, Team, and legacy seat-based Enterprise users receive an automatic 50% Claude Code weekly-limit increase through August 31, 2026 at 11:59 PM PT; Free and consumption-based Enterprise are excluded, and five-hour plus other Claude/Cowork limits are unchangedNot a hard reset, not a five-hour or Cowork boost, and not a permanent plan increase
Z.AISince July 30, new GLM Coding plans use transparent token-based credits; active legacy plans keep their existing calculation through the current cycle, eligible Legacy V2 plans remain renewable/upgradable, and weekends use off-peak treatmentNot a Codex-style hard-reset program and not proof of sustainable unit economics

Anthropic’s promotion terms give a fixed expiry and automatic eligibility rules. Z.AI’s July 30 notice separates new credit metering from legacy five-hour/weekly quotas and preserves a migration incentive. Z.AI’s referral rules separately offer an eligible invited user 10% off the first GLM Coding order.

These are useful contrasts in predictability and accounting. They do not show that Anthropic or Z.AI copied OpenAI, and promotions do not prove that any provider’s economics are sustainable.

Community pulse: a bounded sample, not a survey

The earlier bounded sample covered reactions to the August 8 reset, anticipatory burn around the promised Monday reset, a banked reset expiring early, and uneven reports about whether the five-hour limit returned. The August 20 refresh adds reports of an $8 Plus reset, an $80 Pro 20x reset and shifted recovery date, and uneven rollout plus one reported purchase.

Themes include rapid burn, unpredictable reset value, concern about long-run sustainability, confusion over expiring grants, and speculation that paid resets may eventually replace gifts. This small self-selected sample is not representative. It establishes questions and observed behavior, not provider policy, motive, account-wide prevalence, or a future paid-reset product.

A practical policy without quota burning

Keep a ranked backlog of real quota-heavy work: repository review, research synthesis, migrations, test generation, documentation repair, or a bounded refactor. When capacity is available, pull the highest-value item forward. Keep an urgent-work reserve based on what could reasonably arrive before the displayed recovery time.

If the reserve is threatened, switch routine work to a smaller model where quality remains acceptable. Before purchasing overflow, check the current dashboard for the actual options on the account. Do not manufacture prompts, retries, or subagents merely to drain a meter before a teased reset.

Falsifiable August 20 forecast

AIHackers expects weekly capacity plus account-scoped paid overflow and discretionary promotional resets to remain a plausible near-term equilibrium, without documented universal five-hour enforcement or routine broad banked grants. This forecast fails if OpenAI publishes materially different reset terms, restores dependable five-hour enforcement across accounts, begins routine broad banked grants, or removes the paid-reset control.

For the dated record, use the reset chronology. The Codex claim ledger tracks evidence status. For subscription-versus-API buying decisions, use Smart Spend, the Deals ledger, and the LLM cost-saving playbook.

Frequently asked questions

Is unused subscription capacity the same as purchased tokens?

No. It is a time-limited service allowance, not owned API tokens, cash, credit, or a transferable balance.

Did idling cause the recent Codex resets?

No evidence supports that claim. Provider-wide resets refresh the announced group regardless of one user’s remaining allowance.

Does every reset reach every plan?

No. August 8 and 11 named paid ChatGPT Work and Codex users; August 13 said everyone. The signed-in usage page remains the authority for an account.

Is the Codex five-hour limit active?

OpenAI still documents shared five-hour windows, while current account reports show weekly-only displays. No first-party deprecation or completed-restoration notice resolves the difference, so use the signed-in account as the operational authority.

Should I exhaust my allowance before a teased reset?

No. Prioritize valuable work, retain an urgent-work reserve, and do not burn quota for its own sake.

How is a banked reset different from a hard reset?

An eligible user controls an expiring banked reset. A provider-wide hard reset automatically refreshes the announcement’s group. Account-specific promotional grants can differ.

Is a purchased full reset the same as buying credits?

No. Current community reports describe a purchase that starts a new seven-day limit window and moves the next scheduled recovery date. Credits fund supported work after included limits under a separate ledger.

Where should I check current limits and reset announcements?

Use the signed-in usage page for account state, OpenAI pricing for documented mechanics, and Tibo Sottiaux’s X feed for frequent teases and announcements. The X feed is not a formal policy changelog.

Sources

Last verified: August 20, 2026. Limits, promotions, reset scope, prices, commands, account behavior, and unofficial forecasts can change.