The Codex Reset Fee: A Forensic Analysis of Compute Rationing and Time Arbitrage

The data shows a button that sells time. Not compute. Time. OpenAI’s paid reset function for Codex surfaced in checkout configurations and network resources, excavated publicly by developer Tibor Blaho. Price points: five to eight dollars on the Plus tier. Twenty-five to forty on a tier that has not officially launched anywhere. Fifty to eighty on Pro. Prices exclude tax. The reset does not grant additional quota. It restores quota that once existed. Buy it, and the five-hour window refills. Buy it, and the weekly window refills. The next scheduled weekly reset moves roughly seven days later. That is the entire transaction. Read those mechanics again. The user pays fifty dollars. OpenAI moves a timestamp. Nothing new is created. No additional GPU allocation. No extra inference credits. The user reclaims capacity that was theirs to begin with — capacity spent too quickly for the clock. The system charges a fee for the right to consume one’s own entitlement at a preferred moment. This is not a feature. It is a meter reading with a credit-card form. The most expensive button in Codex manufactures nothing. It re-arranges time. It converts a future allocation into a present allocation at a premium. The premium is the price of urgency. And the price of urgency is the most honest signal a pricing system can emit. OpenAI just published a series of numbers that reveal what it believes urgency is worth. Five dollars. Thirty-two. Eighty. The spread is a map of user value, drawn in cash. Status: UNCONFIRMED. The feature is not in any official changelog. The evidence lives in public checkout configs and developer code digs, not in a press release. But the data is coherent, multi-sourced, and internally consistent across tiers. Treat it as probable, not final. Then treat it as a specimen. CONTEXT: WHAT CODEX ACTUALLY IS Codex is OpenAI’s agentic coding tool. It generates, edits, and executes code inside a sandboxed environment. Subscribers access it through ChatGPT tiers. Plus. Pro. And a strange middle layer that appears in the same leaked configuration: Pro Lite. A tier that has never been announced. The quota structure is two-dimensional. A short window — five hours. A long window — seven days. Each replenishes on independent cycles. Each drains with sustained usage. When both deplete, the tool stops producing useful output. There is no overflow tank. There is no fallback. There is a queue. This is the design detail most commentary will miss. Codex is not software. It is an allocation of remote execution capacity. Every generation step consumes GPU cycles on OpenAI’s infrastructure. Every sandbox execution consumes compute. The product the user subscribes to is not a program. It is an entitlement to a stream of metered compute, rationed through time windows. The subscription is the entrance fee. The quota is the ration. The reset is the legalized black-market bypass — priced, processed, and settled through the same billing system that already handles the purchase of additional credits. OpenAI has not confirmed any of this. No changelog. No help-center article. No executive tweet. The evidence comes from configuration files and network traffic, not from the vendor. A checkout integration exists. Price bounds exist per tier. The Pro Lite designation appears in the same configuration. Either this is a controlled test, a broken deploy, or a roadmap leak. All three outcomes are informative. I have run this pattern before. In 2020, I spent three weeks stress-testing a DeFi lending protocol’s liquidation engine with my own capital. The vulnerability was latency. A fifteen-second price-oracle delay left undercollateralized loans standing when they should have been liquidated. The entire loan book relied on a time assumption: that price discovery would arrive before value left the building. It did not always. Codex’s quota system relies on a similar time assumption. The assumption is that users will stop when the window depletes. The reset feature is OpenAI’s admission that the assumption is false. Users will not stop. They will pay. CORE: THE MECHANICS OF THE METER Call it what it is. A mechanism for converting idle anger into revenue. The reset operates on a simple arbitrage. Between the user’s demand for immediate availability and the network’s demand for load smoothing. The user wants compute now. The network wants demand spread across the cycle. The reset — priced correctly — finds a clearing price for that tension. This is congestion pricing. Same logic as a surge multiplier. Same logic as priority fees on a busy settlement layer. Same logic as paying a premium at a gas station at 2 a.m. The blockchain parallel is exact. Ethereum users pay a base fee for blockspace and an optional priority fee to skip the queue. Codex users pay a base subscription for quota and an optional reset fee to skip the wait. The names are different. The mathematics are identical. THE PRICE LADDER IS A SEGMENTATION MACHINE Five to eight dollars. Twenty-five to forty. Fifty to eighty. The spread is more than tenfold. That is not a single price. It is a set of price points mapped to different willingness-to-pay. Consider the relationship between reset price and subscription price. Plus costs twenty dollars a month. A reset costs a third of that, per event. Pro costs two hundred dollars a month. A reset costs twenty-five to forty percent of that, per event. The heavier user pays more in absolute terms and less in relative terms. That is a loyalty discount dressed as a premium price. It is also demand capture at the top of the curve. The behavioral arithmetic is the real product. A Plus user who resets once a week adds twenty to thirty-two dollars to a monthly bill. That number approaches the cost of a Pro subscription. OpenAI has constructed a pricing ladder that punishes frequent small-user resets and rewards top-tier commitment. This is not an accident. This is a designed migration path. The reset price is the nudge. The subscription tier is the destination. Then there is Pro Lite. Nothing official. Nothing in any marketing deck. But it exists in the checkout config with resets priced at twenty-five to forty dollars. This tier does not exist in public-facing materials. It exists in code. If Pro Lite launches, it will likely sit between Plus and Pro in price and quota. And the reset feature gives OpenAI a direct instrument to measure how much each tier’s users value extra capacity. Before Pro Lite exists, OpenAI can probe demand by observing who hits the reset and how often. The resets are market research that also collects payment. This is not theoretical. In 2021, I analyzed ten thousand transaction records from the Bored Ape Yacht Club floor market. I found that forty percent of observed volume was generated by interconnected wallets. Artificial demand, mechanically manufactured. The lesson was that fabricated activity pollutes every indicator it touches. The reset mechanism has the opposite property. It does not manufacture demand. It reveals it. Every reset click is a clean, self-reported valuation of the user’s own time. A LOAN AGAINST YOUR OWN FUTURE ALLOCATION The evidence states that a full reset restores both the five-hour window and the weekly quota — and that the next weekly reset is deferred by roughly seven days. Pay attention to the second clause. The next reset does not disappear. It is postponed. The user is not buying compute units that did not exist before. They are borrowing capacity from their own next allocation period. Over several cycles, the total weekly entitlement may remain constant. Buy a reset on day ten. The weekly reset moves from day fourteen to day twenty-one. The structural cap — the weekly ration — is not inflated. It is shifted in time. The user is paying to collapse future availability into the present. That is not a feature. That is a time-transformation contract. The product being sold is not compute. It is the option to compress a future time entitlement into the current moment. The premium is the cost of time arbitrage. The user supplies urgency. OpenAI supplies the matching infrastructure. The unit of value is not tokens or credits. It is temporal priority. This structure is the template for metered intelligence. A base fee for a ration. A per-violation fee for urgency. A deferred-compensation mechanism that keeps the ration mathematically constant while making the experience feel unbounded. Yield is just risk wearing a mask of mathematics. In Codex’s case, availability is just cost wearing a mask of convenience. GPU ECONOMICS AND THE PRICE OF CUTTING THE QUEUE Fifty to eighty dollars for a Pro reset. What does that buy in raw compute terms? The industry benchmarks put sustained high-end inference on accelerated hardware at tens of dollars per hour when fully loaded. A heavy agentic coding session — multi-file edits, test execution, long-context reasoning — can consume serious compute. At H100-class economics, an hour of dense inference can approach double-digit costs. A reset that enables several hours of intensive work before the next depletion is not an absurd price. It is a distressed-market clearing price. Here is the part that gets ignored. OpenAI already operates the infrastructure. The marginal cost of fulfilling a reset request is not zero, but it is far lower than the marginal cost of provisioning a second allocation of capacity for every subscriber. The reset monetizes demand spikes during peaks. It can push load into valleys if pricing is tiered by time. It converts frustration into revenue. And if compute supply tightens? Expect inventory control on the reset itself. Daily reset caps. Dynamic pricing. Peak-hour surcharges on top of the reset. I have seen this exact progression in cloud markets. AWS sells Spot Instances at a discount when capacity is idle and terminates them when demand returns. The reset is the inverse instrument. It sells priority when demand is high and lets the window refill when demand is low. It is a two-sided price lever on the same scarce resource. THE UPSELL ARITHMETIC AND THE INVISIBLE TAX Run the numbers again. Plus subscription: twenty dollars per month. A single reset: five to eight dollars. Five resets: up to forty dollars. That is twice the base subscription. The marginal dollar required to sustain an over-consuming usage pattern eventually exceeds the price of the next tier. OpenAI does not need a sales team to upgrade users. They need a usage pattern and a button. This resembles a gym membership with a pay-per-visit surcharge after the sixth visit. Except the gym cannot reprogram your physiology. Codex can reprogram your workflow. Agentic tools consume the five-hour window faster than interactive coding ever did. Each autonomous loop is a draw on the meter. The more autonomous the tool, the more frequently the window depletes. The more frequently the window depletes, the more frequently the reset becomes attractive. There is an accounting signal buried in the evidence. The phrase “excluding tax” appears in the data points. That phrase is a tell. It indicates an invoice. An invoice implies reimbursement. Reimbursement implies a business buyer. A developer with a corporate card does not feel an eighty-dollar reset. The purchasing decision is detached from personal income. That is the highest-margin segment in any pricing table. I reviewed the custodial and settlement infrastructure of spot Bitcoin ETF applications in 2024. The conclusion of that work was blunt: institutional entry does not eliminate operational risk. It shifts it. OpenAI’s pricing carries the same signature. Institutional adoption of Codex will not eliminate compute scarcity. It will shift it into procurement. A reset becomes a line item. An operations manager bakes it into the budget. The cost becomes an entitlement. The entitlement becomes dependency. The dependency becomes recurring revenue. HARVESTING WILLINGNESS-TO-PAY CURVES This is the dimension most commentary will miss. The reset is a price-discovery instrument. Every time a user clicks reset, OpenAI learns something. The frequency of resets per tier estimates the demand curve for additional capacity. The tier bounds the user’s time value. A Plus user who resets at five dollars reveals that their marginal coding hour is worth at least five dollars. A Pro user who resets at eighty dollars reveals a far higher valuation. Aggregated across the user base, that data maps the distribution of urgency in the market. That data feeds the next pricing experiment. It informs GPU procurement. It informs tier segmentation. It tells OpenAI where to deploy capacity and when to raise prices. The reset is not merely a product. It is a sensor network. Every purchase is a data point. Every abstention is a data point. The system can now distinguish the price-sensitive from the time-sensitive with precision. This is the same methodology I applied to the 2022 Terra collapse. I reconstructed the withdrawal flows across five centralized exchanges and estimated that a hundred million dollars in Anchor Protocol withdrawals was sufficient to trigger the death spiral. The thesis was binary. The mechanism was fragile. The yield was a fixed promise on a non-productive asset, and fixed promises on non-productive assets are always broken. Codex’s reset has no such fragility. But it has a different kind of hidden dependency. The dependency is on users accepting a metered model at all. The moment the market rejects metering, the narrative flips. Not because the mechanism is wrong. Because the expectation was subscription-grade predictability. The reset is the wedge that breaks that expectation. Once a user pays for urgency, the subscription is no longer the product. The meter is the product. THE RATE LIMIT AS ARCHITECTURE Look deeper at the dual-window structure. Five hours. Seven days. Two independent clocks. Why both? The short window smooths adversarial or accidental bursts. A user cannot drain a month of compute in a single afternoon. The long window caps sustained consumption. A user cannot run an autonomous fleet of coding agents around the clock without hitting a hard ceiling. Together, the two windows form a rate-limiting grid that protects the shared resource pool from both spikes and drains. The reset attacks that grid at its weakest point: the boundary between the two windows. Restore the short window, and the user can resume immediate work. Defer the long window, and the sustained cap holds. The design is careful. It does not double the monthly ration. It front-loads it. The architecture is a textbook example of load management. The pricing is the enforcement mechanism. Silence in the logs is louder than the crash. OpenAI has not denied the feature. It has not confirmed it. That silence is a deliberate state. The company is watching how the leak propagates, how developers react, what the community frames as acceptable. The leaked config is a free public focus group. The absence of a statement is the statement. COMPETITIVE DYNAMICS: THE UNLIMITED LIE The current market for AI coding tools is dominated by fixed monthly subscriptions. GitHub Copilot. Cursor. Claude Code. The marketing language is familiar: unlimited usage, Pro access, all-included. That language is fiction. Every provider rations capacity. Some enforce soft limits. Some throttle silently. Some degrade quality under load. The difference between them and OpenAI is that OpenAI is testing a disclosed price for exceeding the ration. The disclosed price is the differentiator. It converts a hidden operational decision into a visible economic one. When a competitor says “unlimited,” they are either lying, rationing silently, or pricing access high enough that only a fraction of users consume enough to matter. The reset is the anti-“unlimited.” It declares scarcity openly and prices access to it. Expect competitors to follow. Not with “unlimited resets” — that would be financial suicide — but with their own tiered-overage structures. The race is not toward free. The race is toward honest metering. The provider that prices scarcity accurately wins the trust of power users. The provider that hides scarcity loses them when the meter finally appears. The DeFi parallel is unavoidable. Every new chain slices liquidity rather than scaling it. Every new “unlimited” coding plan slices trust rather than providing capacity. This is not scaling. It is fragmentation dressed as generosity. CONTRARIAN: WHAT THE BULLS GOT RIGHT The easy take is that OpenAI is gouging its most loyal users. That take is lazy. Consider the alternative. An “unlimited” tier requires OpenAI to provision infinite capacity at a fixed price. No one does that. Not in cloud compute. Not in DeFi. Not in any scarcity market. When a competitor markets “unlimited,” they are either hiding the ration or pricing it so high that the average user never discovers the limit. The reset removes the deception. It says: here is the meter, here is the price of skipping the queue, here is the cost of now. That honesty has real value. Users who value their time can express that value. Users who can wait can wait for free. The reset adds one option to a choice set that previously had one item: wait. No amount of urgency could change the outcome before. Now urgency has a price. That is not exploitation. That is efficiency. The reset also creates the possibility of self-selection. The user who resets at eighty dollars is revealed as a high-value customer who deserves premium support, capacity guarantees, and service-level commitments. The user who never resets is revealed as price-sensitive and can be served with standard capacity. The system segments itself. That reduces friction for everyone. There is a deeper point. The reset is a test of the paid-overlay model for AI itself. If it succeeds, expect metered AI to become standard across the industry. If it fails — if backlash forces OpenAI to withdraw it — the signal is equally valuable. It tells the market where the tolerance threshold for post-subscription fees sits. That threshold is market infrastructure. Knowledge with a price tag. In 2018, I spent six weeks manually auditing a smart contract codebase and found a critical reentrancy vulnerability in a token swap function. The fix was a mutex — a lock — preventing a function from being re-entered before the first execution completed. The principle was that a system’s trust must not depend on goodwill. It must survive adversarial pressure. Codex’s quota system is a kind of lock. It holds users to a rate limit. The reset is the system’s admission that locks can be breached by legitimate force — a purchase — rather than by a bug. Treating the user as a paying adversary with a budget is harsh. It is also honest. There is a transparency in the reset that most pricing mechanisms lack. The bulls are right about one more thing. The reset is a measured, incremental step. It does not change the core subscription. It does not remove the free waiting path. It does not penalize users who plan around the quota. It adds an optional premium lane. That is the most rational entry point for a pricing evolution. Start with the optional lane. Measure the take rate. Adjust. Scale. OpenAI is running a controlled experiment, not a policy mandate. THE CRYPTO CONNECTION: COMPUTE AS A SETTLEMENT LAYER The deeper structural story is not about OpenAI. It is about the commoditization of intelligence and the settlement layer that will price it. AI agents do not have credit cards. When autonomous coding agents manage their own compute budgets — when they decide whether to wait for a quota window or pay a reset premium — the pricing mechanism must be machine-readable. A button in a checkout config does not scale. A programmatic meter, an API for purchasing priority, a settlement rail that agents can transact on — that is the next layer. This is where crypto stops being an analogy and becomes infrastructure. Machine-payable compute requires cryptographic identity, programmatic payments, and transparent pricing. The reset button is the prototype. A centralized prototype, priced in fiat, settled by OpenAI. But the design pattern is exactly what decentralized compute networks have been attempting for years: metered access to scarce hardware, priced by urgency, consumed by software. The floor is an illusion. The floor is a trap. That was true for NFT floors in 2021. It was true for the Terra peg in 2022. It is true for the subscription floor of the AI coding market. The flat fee is not the product. The product is the ability to consume a scarce resource on demand. And the market has just been told what that consumption costs when the queue is long. The meter is running. TAKEAWAY: THE PRICE OF NOW The arithmetic of the reset is simple. The implications are not. Watch the data. If resets cluster at the low end of the price range, demand is thin and OpenAI will adjust. If resets cluster at the high end, the market is declaring that professional coding time is worth more than a hundred dollars an hour in marginal compute premiums. If Pro Lite launches publicly, the tier structure is being rebuilt for the metered era. If the reset disappears without comment, the experiment failed internally. Silence in the logs is louder than the crash. Every tool that sells intelligence will eventually face the same question. What is the price of now? Codex has answered with a number. Five dollars. Eighty dollars. The number depends on who you are. The system records the answer every time. Precision is the only currency that never inflates. The resets are precisely priced. The open question is whether users will remain precisely informed. The meter is not the enemy. The hidden meter is. OpenAI’s leak is a gift. It lets the market see the pricing architecture before it is officially deployed. That is rare. Treat it as evidence. Treat it as a warning. And read the code next time before you buy the promise.