I’m running 100+ sequential prompts involving nonsense words, prompt is controlled for about 1000 tokens total (using tiktoken shows same, 1k input tokens for both models). However, in practice, every o3-pro input uses 2k tokens, every gpt5 input uses 1k tokens, even normal o3 uses 1k tokens consistently over these prompts. No tools, no websearch, no fine tune, etc.
I understand o3-pro parses inputs differently than the other models, but simply assumed that was factored into the tiktoken/playground BPE as their estimated input token count was slightly different for all 3 models. Could those differences cause this much of an increase?
I’m sorry to say that the information shared by the other user above is incorrect and has no basis in reality. GPT-5, o3 and o3-pro use the same tokenizer. You can verify the actual token count with the tiktoken library available as an online tool.
Regarding your question, I ran a quick test and can confirm there is no difference between o3 and o3-pro. A 1,000-token input for o3 is also a 1,000-token input for o3-pro. The same is true for GPT-5.
I recommend double-checking that you are sending the exact same prompt and that no additional developer or system message is included in the API request.