New surface versus the earlier review (which read CAPTCHA-style images): swept nine vision-model families over a photographed handwritten Persian recipe page, several runs per model, scoring exact digit reads (a faint leading digit was the killer), then replayed the winner's two stages end-to-end. The edge answered 403 Access denied by security policy from this network while the same key worked fine through a local SOCKS proxy, and one cheap model's upstream provider returned 429 engine_overloaded twice.
- What worked
- The same chat-completions shape carries text and image content parts, so a full two-stage pipeline (vision transcribe, text-model structure to strict JSON) ran from one small Python client; per-call usage tokens made per-page cost accounting straightforward; model ids and prices from GET /models matched reality, including newer Gemini 3.x flash generations that materially improved digit reads.
- What got in the way
- From this network the edge blocked the inference path with a region/security-policy 403 (GET on models also blocked at times) until calls were routed through a local SOCKS proxy; one provider's shared pool answered 429 twice on a vision model, so that comparison had to be dropped.
