When an AI Vendor's Bug Spoils Client Work, Who Pays?

On September 25, 2026, OpenAI fixed a bug that had degraded how GPT-6 Sol and Luna read images, and recommended that customers re-run affected evaluations and retry affected workflows. For an agency that had already delivered image-based work to a client, the re-run is the easy part. In the worked example below, the model costs about $22 to run again. The people who check the results cost between $580 and $6,580.

That gap is the whole problem. Most agency contracts say nothing about a model provider’s defect, so the cost falls wherever the pricing model happens to put it. This post prices one illustrative job, shows who pays under three common pricing models, and suggests a clause that settles the question before the next bug.

The job is invented to make the arithmetic concrete; every volume, time and rate below is illustrative, and only the model prices are real. An agency extracts supplier, date, amount and tax fields from 4,000 scanned invoices for a client, using GPT-6 Sol in the days after its September 22 launch. It delivers the data to the client’s finance system. On September 25, OpenAI’s changelog reports the image fix and advises re-running affected work.

Nobody yet knows whether the delivered data is wrong. OpenAI did not say how badly image reading was affected, as we noted in our post on the fix. The responsible course is to re-run, compare and check. The question is what that costs and whose budget it comes from.

OpenAI’s GPT-6 Sol model page lists $2 per million input tokens and $10 per million output tokens, with Batch at half those rates. We assume 3,000 input tokens per invoice for the image and instructions and 500 output tokens for the extracted fields. That is $0.011 an invoice at standard rates, or $44 for the job, and $22 on Batch, which suits a re-run with no deadline.

If the sample is clean, the total is $602, of which the model is about 4%. If the sample finds errors and every invoice is reviewed, the total is $6,602, and the model is about a third of 1%. The same pattern appears whenever AI output needs checking, which we worked through in when checking AI output costs more than generating it. A vendor bug simply makes you pay the checking cost twice.

None of the three is fair by default. Under a fixed fee, a clean sample costs the agency a tolerable $602; a full review wipes out $6,602 of margin on a job that may have been priced at a few times that. Under hourly billing, the agency is covered on paper, but asking a client to pay for a model provider’s fault is a hard conversation. The changelog entry does not mention credits or refunds, so neither party should count on recovering the cost from the vendor.

The pricing models themselves are covered in our AI agency pricing guide. What none of them includes by default is a rule for this case.

The fourth point is the one that makes the others workable. GPT-6 Sol has no dated snapshot to pin, so the model name alone does not tell you which version produced a file. A dated log of runs does. Without it, an agency cannot say which deliverables fall inside the window, and the default becomes reviewing everything.

The clause also helps with pricing. An agency that knows its exposure can price a small reserve into fixed-fee AI work, rather than discovering the risk the week a bug is announced. Our statement-of-work framework shows where terms like these sit in a contract.

The worked example uses a 5% sample, 200 of the 4,000 invoices, and that number was chosen for the arithmetic rather than handed down by anyone. What the sample is for is more useful than its size. It is the cheapest piece of evidence that can tell you whether the full review is needed, so it should be large enough that a clean result means something and small enough that you will actually run it the day the vendor’s notice arrives.

A clean sample lets the agency say one thing to the client: on this many re-run invoices, the corrected model produced the same fields as the delivered data, so the delivered data is probably right. It does not prove every invoice is correct, and the note should not claim that it does. A sample that finds errors changes the question from whether to review to how much, and the clause in the previous section decides who pays for the answer.

Agree the size in the contract, not on the day. An agency that proposes a 5% sample after the bug is announced looks like it is choosing the cheapest option; one whose statement of work already says “a sample of N items, reviewed by the agency at no charge” is following a rule both sides signed. The same clause should say what finding rate triggers the full review, so that too is a rule rather than a negotiation.

One more number belongs in the same clause: how long the agency keeps the run log. A vendor’s notice can arrive after the work has shipped and been invoiced, and a log that was pruned after the invoice cannot answer which deliverables fall inside the window. Keeping it for the cover window plus a margin is a small storage cost against a large review one.

The illustrative cases below are the conversations the clause is written to prevent. In each, both parties behave reasonably and one of them ends up paying for something they did not cause.

The missing log: nobody recorded which model and date produced each deliverable, so when a cover window applies the agency cannot say which jobs fall inside it, and the only defensible option is to review everything. The promised refund: the agency tells the client the vendor will cover the cost, when the changelog entry mentions neither credits nor refunds, and ends up covering it from margin after making the promise. The surprise invoice: an hourly agency bills the review hours as usual and the client, learning of the bug from the invoice, refuses to pay for a defect neither party caused. The sample read as proof: a clean 5% sample is reported to the client as confirmation that the delivered data is correct, and an error found months later turns a defensible process into an apparent misstatement.

The fourth point of the clause, logging model and date per deliverable, prevents the first. The clause’s second point prevents the third, because the client knew in advance which review hours could be billed. Honest wording prevents the second and fourth, and costs nothing.