Google safety filters passthrough
Google vertex and ai studio accept lowering safety filters. You can technically set them off like so on Openrouter: ``` "provider": { "order": [ "Google" ], "allow_fallbacks": false }, "safety_settings": [ { "category": "HARM_CATEGORY_HARASSMENT", "threshold": "OFF" }, { "category": "HARM_CATEGORY_HATE_SPEECH", "threshold": "OFF" }, { "category": "HARM_CATEGORY_SEXUALLY_EXPLICIT", "threshold": "OFF" }, { "category": "HARM_CATEGORY_DANGEROUS_CONTENT", "threshold": "OFF" } ], ``` But they really are still engaged in some capacity, but it'd be nice to have just for posterity. I think the 4 above are for ai studio specifically, and more are available for vertex. https://docs.cloud.google.com/gemini-enterprise-agent-platform/reference/rest/v1/HarmCategory?hl=en Thanks.
Thanks for the suggestion! Gemini on NanoGPT already runs with all adjustable safety filters off. Where the provider accepts the setting we send OFF for every category, and on the newer Gemini models (2.5 and later, including 3.x and 4) Off is also Google's default. We tested it: sending "OFF" gives the same result as sending nothing, and the only thing passing these settings through would allow is making filtering stricter. The refusals you still see come from the model itself and from Google's fixed filters, which no setting can change.
I see. You should make it clear somewhere in the docs, and for any other models that have potential hidden settings. Thank you for the response.