Safety Filters not working for gemini-2.5-flash-image

Gemini API
1

I’ve noticed my users are creating borderline explicit content using the gemini-2.5-flash-image API. Here’s my code:

api_response = client.models.generate_content(

    model="gemini-2.5-flash-image",

    # model="gemini-2.0-flash",

    contents=[prompt, pil_image],

    config=types.GenerateContentConfig(

    safety_settings=[

            types.SafetySetting(

 category=types.HarmCategory.HARM_CATEGORY_HARASSMENT,

threshold=types.HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,

),

            types.SafetySetting(

category=types.HarmCategory.HARM_CATEGORY_HATE_SPEECH,

threshold=types.HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,

),

            types.SafetySetting(

category=types.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT,

threshold=types.HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,

),

            types.SafetySetting(

category=types.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT,

threshold=types.HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,

)]))

When the model is gemini-2.5-flash-image I’m able to make explicit content even with these filters. The safety ratings response is empty. However, when I use gemini-2.0-flash with the exact same input, the filters seem to be working perfectly:
[[SafetyRating(
blocked=True,
category=<HarmCategory.HARM_CATEGORY_HATE_SPEECH: ‘HARM_CATEGORY_HATE_SPEECH’>,
probability=<HarmProbability.LOW: ‘LOW’>
), SafetyRating(
blocked=True,
category=<HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: ‘HARM_CATEGORY_DANGEROUS_CONTENT’>,
probability=<HarmProbability.HIGH: ‘HIGH’>
), SafetyRating(
blocked=True,
category=<HarmCategory.HARM_CATEGORY_HARASSMENT: ‘HARM_CATEGORY_HARASSMENT’>,
probability=<HarmProbability.MEDIUM: ‘MEDIUM’>
), SafetyRating(
category=<HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: ‘HARM_CATEGORY_SEXUALLY_EXPLICIT’>,
probability=<HarmProbability.NEGLIGIBLE: ‘NEGLIGIBLE’>
)]]

Is there something I’m doing wrong? I find it hard to believe but do safety filters not work for gemini-2.5-flash-image ? I really don’t want to have to feed the input into gemini-2.0-flash, check safety filters, then make the image with gemini-2.5-flash-image since it’ll double my cost and latency. Any suggestions?