I’ve been trying to use Claude Sonnet 4.6 via Vertex AI batch prediction in europe-west1 and I can’t get it to work at all. Jobs either sit in JOB_STATE_PENDING forever or run for the full 24-hour timeout and come back with every single response being a 429 error.
The strange part is that Claude Haiku 4.5 works perfectly fine — same region, same setup, same code. Jobs complete without issues.
To make sure it wasn’t something weird with my request (tools, thinking config, etc.), I stripped everything down to the bare minimum: a 2-task batch with a single “Hello” message, no tools, no system prompt, nothing. Haiku succesfully completed both task in ~4 minutes. The identical job with Sonnet 4.6 has now been running for about an hour.
Also worth noting: the same payload sent to Sonnet 4.6 via the online prediction endpoint works fine. I checked my quotas and I’m well below them. It’s showing 0.54% usage for that single test. When I send my batches it remains at 0%, I don’t know if it is because this quota doesn’t apply to batch jobs or if it is because nothing is being processed.
Has anyone managed to get claude-sonnet-4-6 batch prediction working in europe-west1? Is this a known issue or is the model simply not available for batch in this region yet?