Can confirm. This started for us about 1 day ago, in all streaming code that has been deployed just fine since probably January.
See Malformed streaming answers from GPT-4 completions API lately
The first chunk is unusual in that it has two 'data: ’ payloads separated by a newline (which is fine, though I am not sure we ever saw that in practice), but the second payload is only half the completion object. The second chunk is the rest of the object, but is invalid because it does not start with 'data: '. Then the rest of the stream is fine: a single 'data: ’ payload in each chunk