One /goal invocation. A single, misdeclared objective. In the worst observed case, it consumed between 15% and 70% of a paid OpenAI Codex user's weekly usage limit before terminating. That is not a rate-limit snag. That is a structural burn fault. On March 6, OpenAI confirmed the diagnosis: eight discrete issues were inflating consumption across Codex and ChatGPT Work. They patched them, and reset usage limits for all paid accounts. Depending on how you work, the same allocation now lasts 10–50% longer. This is not a courtesy. This is a critical audit finding.
I have spent most of my career auditing systems where hidden leaks cause downstream failures. In 2026, I tracked 5,000 AI-driven wallets on Solana for three months. The most common failure was not malicious attacks—it was agents running in endless loops, paying gas, returning nothing. Codex's weekly limit is a private ledger of AI labor. When it drains without output, you are not losing tokens. You are losing yield. The reset is real money, but only if the underlying burn is fixed.
Let's enter the audit trail. The first bug class is /goal. This is the most dangerous because it is a termination-condition failure. A user sets an objective. The agent executes. The agent finishes. Then it keeps going. In a smart contract, that is a missing halting condition—a DoS vector. In Codex, it is a quota killer. One bad /goal run can consume up to 70% of a weekly allocation. That is not a rounding error. It is a systemic absence of a task_complete flag.
The second bug class is Computer History. Every session, the agent summarizes prior operations. The summary itself is useful. But in certain configurations, that summarization consumes roughly 20% of the weekly limit. Think of it as replaying an entire ledger just to check the final balance. The information gain is marginal. The cost is massive. A proper design would cache historical context locally, not re-derive it every session.
Additional issues were found in image compression, background memory, automated tasks, sub-agents, and MCP. Each is a different failure mode, but they share a common root: the agent does not account for the cost of its own internal operations. Image compression might seem trivial, but when an agent recompresses the same asset repeatedly, the token cost compounds. Background memory behaves similarly—constant recall of old sessions adds hidden load. Automated tasks and sub-agents multiply the problem because each new process brings its own overhead. MCP, the Model Context Protocol layer, adds another boundary where token accounting can go wrong.
Here is the core insight: AI agents do not have a built-in gas meter. In crypto, every operation has a fee. In Codex, the fee is opaque, aggregated, and only visible after the damage is done. The eight bugs are not independent incidents. They are symptoms of a missing metering architecture. If the agent cannot see the cost of an action before taking it, it will waste resources. This is why the fix period matters. OpenAI says the same limit now lasts 10–50% longer. That is an admission that prior burn rates were not just user error. They were protocol waste.
Based on my audit experience, I would add a ninth issue that OpenAI did not mention: observability. Users have no real-time dashboard showing where the weekly limit goes. Without that, every bug fixes is a blind patch. You cannot optimize what you cannot measure. The reset is a bandage. The longer-lasting fix is a per-action cost readout. Until then, the user is flying with a fuel gauge that only works at zero.
Now the contrarian angle. When OpenAI resets usage limits, the market treats it as a gift. It is not. It is a liquidity injection. The same playbook has run in DeFi for years: offer inflated incentives, attract deposits, then quietly reduce subsidies once behavior locks in. Yields attract capital; sustainability retains it. Codex resets are no different. Every reset buys time for users to forget the burn. But the underlying architecture still lacks disciplined cost control. If you are a paid user, you are not a customer. You are a beta tester with a wallet.
Trust is a variable, not a constant. The data tells us that /goal was consuming up to 70% of a weekly cap. That is not a bug found in testing. That is a bug shipped to production and paid for by users. The reset is necessary, but it does not restore trust. It only resets the counter. The platform must demonstrate, with data, that the same task now costs meaningfully less. Anecdotal promises from Tibo—"10 to 50% longer"—are not a confidence interval. I need a before-and-after chart, not a press release.
This weekend may bring another reset. Yesterday, Tibo hinted that Codex is approaching a user milestone and suggested a celebration. After today's early reset, he said the button was already pressed, so the celebration moves to tomorrow. Historical precedent supports this: previous resets followed milestones of 15 million and 20 million users. If Codex has crossed the next threshold, another reset is likely. Do not mistake that for generosity. It is engagement farming. Volatility is the price of permissionless entry, but predictable resets are a marketing strategy.
The forward-looking question is simple: will your limit actually last longer next week? If yes, the patches were structural. If no, the reset was just a morale operation. Watch the burn rate per completed task. Track the session cost before and after the fix. That is the only evidence that matters. The exit liquidity here is not someone else's entry error—it is your weekly allocation spent on an agent that never learned when to stop.
The button was pressed. The caps are refilled. The real audit begins after the celebration.