This has been extremely common with the closed source providers.
Back in the GPT3 era OpenAI would release a new model, people absolutely loved it. It would be very uncensored and they'd go nuts. Then once the press and benchmarks were in they'd turn the censorship back on, crippling the model but then it was "safe" again.
And then of course providers routing some queries to smaller models if they deem it "dangerous" (or need the efficiency probably).
Open models with third party trusted providers tend to at least give you what your asking for. Although in that case you also need to double check what quantization level its using, but that data tends to be transparent.