AI adoption
OpenAI’s API changelog is the page where the provider records, day by day, what changes for businesses that call its models. Its entry for 25 September 2026 states that a bug in image encoding degraded the visual understanding of GPT-6 Sol and GPT-6 Luna, two models released on 22 September, and that the fix improves results on visual tasks in the API and in Codex, including when an agent operates a computer. OpenAI recommends that customers whose use cases involve images rerun their evaluations and reprocess affected workloads. No start date for the bug is given, and no measure of its scope either. The same changelog shows, on 13 March 2026, a comparable fix to the GPT-5.4 image encoder, presented at the time as requiring no action. Two fixes to the same component in six months, with two different instructions.
Our perspective
A model called through an API is not a fixed component: its quality can drop and then recover without the customer being told, and the provider decides whether action is needed. A mid-sized company that has a model read invoices, drawings or site photos has only one way to know: its own evaluation set, rerun at every change. Our advice: build fifty reference cases per visual use case, with the expected answer, and rerun them at every new entry in the provider’s changelog.
Self-interested source: the provider describes its own bug, without dating its start or measuring its scope.
Sources
OpenAI, API changelog, entries of 25 September 2026 and 13 March 2026, accessed on 28 September 2026: https://developers.openai.com/api/docs/changelog
Explore further
The Spentia newsletter
Three dated, sourced facts, an article and a podcast. Delivered in French every Monday.