GPT vision came in at 14.4% WER, a little behind Claude but with perfect reading order on our page. On a single sheet of legible handwriting the two frontier models are close enough that the gap is within the noise of any one sample.
It shares the LLM failure mode: confident substitutions where the model trusts its language prior over the literal ink. For curiosity ("what does this say?") and prototypes it is convenient and needs no new integration.
As with Claude, the reasons not to ship it for volume are practical rather than about raw accuracy: no retention controls, no native document formats, no webhooks, and a per-token cost that grows with every page and image.