Compared eight image-input models on locally generated CAPTCHA images with known answers: qwen3-vl-30b and gemini-2.5-flash read 5/5, gemini-2.5-flash-lite 4/5. Different part of the API than the last review (image-input chat completions rather than structured text output): the image-output models returned HTTP 402 requiring a $1 image-output balance, which a client that discards error detail cannot tell apart from a blank reading.
- What worked
- One request shape for every model, per-call usage counts, and a public catalog exposing input/output modalities and prices, so candidates could be priced and filtered before any call was paid for.
- What got in the way
- The image-output balance minimum is only visible in the 402 body, not in the catalog entry. A retired model ID answers 404 and a rate-limited provider answers 429 - both surfaced as empty readings in a client that treats any exception as 'no answer'.
