Infrastructure

OpenAI Paid You Back in the Thing That Broke

Four incidents on 27 July alone, on OpenAI's own status page. Compensation for the weekend outage was a usage reset, which is credit on the service that failed.

InfrastructureBy Servola Tech Desk2026-07-274 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

OpenAI Paid You Back in the Thing That Broke

Key takeaways

OpenAI's own status page logged four separate incidents on 27 July, and independent tracking of that history counts roughly 166 incidents across about nine months, an average near 18 a month. The published record is the vendor reliability document almost nobody reads before signing.
Compensation for the weekend disruption came as usage-limit resets for Codex and ChatGPT Work users, which is credit denominated in the same service that failed rather than money or a contractual service credit.
The reset destroyed value for users who had spent a reset shortly before it arrived, because a reset restores an allowance rather than adding to a banked one. A remedy that overwrites is not a remedy.
Check which door your organisation came in through. The remedy on a seat-based subscription and the remedy in an API agreement are rarely the same, and only one of them is usually written into something you can enforce.

Four incidents in one day, published by the vendor

On 27 July, OpenAI's status page recorded four separate incidents. Elevated errors affecting ChatGPT conversations at 4:32pm. Elevated latency, timeouts and interrupted streaming on gpt-5.1 mini and gpt-4.1 mini across the API at 6:06pm. Image generation unavailable in ChatGPT at 7:04pm, and unavailable again at 8:24pm. Each was closed with the same sentence: all impacted services have now fully recovered.

That was not an unusual day. The 25th carried two incidents of elevated error rates. The 24th carried three, including elevated API error rates and latency on the gpt-image-2 model and elevated errors in Codex Review. Saturday's disruption was the visible one, touching fifteen ChatGPT components, four Codex components and twelve API components, with users meeting 503 responses that carried an internal circuit-breaker label. OpenAI moved from investigating to monitoring within about an hour.

Independent tracking of the same status history counts roughly 166 incidents across about nine months, an average close to 18 a month. None of that is leaked or inferred. It is the vendor's own published record, updated in real time, free to read, and almost never read by the people signing the contract.

The remedy was denominated in the broken currency

OpenAI's response to the weekend disruption was to reset usage limits for Codex and ChatGPT Work users. Read that as an instrument rather than as a gesture. The compensation for a service being unavailable was more allowance on the service that had been unavailable. It is not money, it is not a contractual service credit, and it cannot be spent anywhere else.

Then came the genuinely instructive part. Users who had already consumed one of their resets shortly before the compensation landed found it did nothing for them. A reset restores an allowance to full. It does not add to an allowance already banked. For that group the goodwill gesture arrived as a write-off of value they were holding, and the complaint was not ingratitude. It was arithmetic.

This failure mode is worth naming, because it will recur across every consumption-priced product. When the unit of compensation is the same unit that was disrupted, the remedy's value depends entirely on the customer's position at the moment it is granted. Some customers are made whole. Some are made worse off. Nobody is paid.

Eighteen a month is a number, not a feeling

For most buyers the useful move here is unglamorous. Take the incident count from the vendor's status history, write it into the file next to the marketing availability figure, and note that the two describe different things. An availability percentage is an aggregate over time. An incident count is a frequency of disruption, and frequency is what actually interrupts your staff.

If your organisation is a financial entity inside the scope of the European digital operational resilience framework, this is already a formal obligation rather than good practice. Information and communication technology third-party risk has to be registered, assessed and monitored, and a provider that publishes its own incident history has effectively handed you the evidence. The awkward part is that the evidence sat in public the whole time.

The wider point holds outside regulated sectors too. Any dependency a team touches dozens of times a day deserves a measured reliability figure rather than an impression. Impressions are formed by the worst outage anyone remembers. A status history is formed by all of them.

Three questions to answer before Friday

First, which agreement governs your usage. A seat-based subscription, an enterprise agreement and an API contract typically carry different commitments, and the remedy for downtime in each is rarely the same. Establish whether your organisation is entitled to a service credit, to a refund, or to nothing beyond whatever the vendor decides to grant.

Second, what a disruption costs you per hour. Not in theory, but with the number of people affected and the work that stops. Firms routinely discover at renewal that they never calculated it, which is precisely why a remedy denominated in tokens was accepted without anyone noticing what it was.

Third, what happens to the work in flight. Saturday's reports were dominated by developers who lost momentum rather than data, but the same disruption pattern reaches document workflows, customer-facing assistants and anything with a queue behind it. Decide in advance whether a two-hour interruption is an inconvenience your team absorbs or an incident your customers see, because that determination changes which vendor commitments are worth paying for.

Frequently asked questions

Why is a usage reset not the same as a service credit?

A service credit is a contractual entitlement that reduces what you pay, and it is enforceable because it sits in an agreement. A usage-limit reset is a discretionary restoration of allowance on the service itself. It carries no monetary value outside the product, it cannot be claimed if the vendor chooses not to grant it, and its worth to any given customer depends on how much allowance that customer happened to be holding when it arrived.

How do I find a vendor's real incident history?

Most large providers publish a status page with an incident archive, usually available without an account. Count the incidents over a fixed window rather than reading the latest entry, because the useful figure is frequency over months rather than the state right now. OpenAI's own history records four incidents on 27 July alone, and independent counts put roughly 166 across about nine months.

Does this matter if we are not in a regulated sector?

The formal obligation applies to financial entities under the European digital operational resilience framework, but the underlying exposure does not check your sector. Any tool used dozens of times a day by a team is an operational dependency, and the questions are identical: how often does it fail, what does an hour cost you, and what are you contractually entitled to when it does.

There is no scandal in a fast-growing service having incidents, and publishing them in the open is more than several competitors manage. The problem sits on the buyer's side of the table. The reliability record was public, the remedy structure was knowable, and most organisations discovered both on the Saturday it mattered. Pull the status history for every vendor your business runs on and put the number in writing this week.

OpenAIChatGPTCodexservice availabilityvendor riskDORAstatus pageservice credits

Nvidia May Guarantee 250 Billion of OpenAI's Rent

OpenAI holds no investment-grade rating, so Nvidia is reportedly guaranteeing the lease on a 10 gigawatt Ohio campus. A chipmaker's credit is now a line in your compute budget.

4 min read

Gaming

The Disc in the Tray Needed a Server

Xbox's outage blocked disc-based and backwards-compatible games because the licence check failed. What a launch-time phone-home costs you under EU rules.

5 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles