Claude vision returned 11.2% WER on our sample, competitive with the cloud document APIs and slightly ahead of GPT-5. It reads layout well and, in our longer tests, held together marginally better than other vision models across multiple pages.
The failure mode to understand is silent correction. Even with a strict "transcribe exactly" instruction, the model occasionally replaces a word it is unsure about with a plausible neighbour. That is more dangerous than an obvious error because nothing flags it, which matters when you cannot proofread every page.
For a prototype, a one-off, or a workflow where a human checks the output, it is genuinely useful and requires no new vendor. For unattended document volume, the lack of retention controls, native export formats, and webhooks means you end up rebuilding a document pipeline around a per-token bill.