Provisioned Rate exceeded on DeleteFaces in Rekognition

I am getting a Provisioned Rate exceeded error when calling the DeleteFaces api of AWS rekognition, using the current version of the .net DSK.

I think what triggered the error in the first place is that I was running some batches in parallel, and even though they don't make many calls to AWS rekognition, it happened occasionally that their calls were simultaneous, exceding the limit of 5 per seconds (I managed retry attempts but I had not realised the SDK had its own retry policy by default).

However it seems now that any single call I make to DeleteFaces returns a "Provisioned Rate exceeded" exception, no matter how long I wait before I retry (>1h), or how many face ids I try to delete per call (50 to 4096), or the retry policy of the sdk (legacy or standard). Other functions of the rekognition API seem to work fine. Is there a rate limit other than the 5 calls per seconds advertised as a default for rekognition (and when that limit is reached, how long does it take before it allows new calls)? Or is it some sort of shadow ban after a certain number of provisioned rate exceeded errors (and if so how long should I wait before the ban is lifted)? Does it matter how many faces I delete in one call? I already batch faceids by batches of 4096. Do these get further split by the .net sdk in multiple calls, resulting in some sort of api calls amplification?