It’s very content specific, what are you processing with the API?
One of my little side projects right now is translating Russian fiction, specifically a genre over there called ‘boyar-anime’ which is essentially fantasy set in imperial russia. I do most my heavy translation using Anthropic Haiku which is very cheap and unlike the higher end models it tends to dumb down some of the more complex parts of Imperial Russian aristocracy so it’s more in line with similar fiction over here. When I take the source book, I chunk it down into small segments that I translate individually so I don’t get context bleed, then I mechanically process to find anything that didn’t translate very well. I combine roughly 40 of these weirdly translated segments into a jsonl file and submit the file through the API. OpenAI Batch API can accept up to 900k tokens, but you’ll wait close to 11 hours for something that large. 40 segments is around 30k tokens and that usually processes in a few mins to an hour depending.
The jsonl file is essentially made up of smaller json blocks
{"custom_id":"SEGMENT-NUM","method":"POST","url":"/v1/responses","body":{"model":"gpt-5.3","input":[{"role":"system","content":[{"type":"input_text","text":"You are a meticulous English language proofreader."}]},{"role":"user","content":[{"type":"input_text","text":"PROMPT - SUBMITTED SEGMENT"}]}],"max_output_tokens":8192}}
I then setup polling to check back with the API every few mins, when the submitted queries are completed, I send more automatically until everything has been processed.
It’s very content specific, what are you processing with the API?
One of my little side projects right now is translating Russian fiction, specifically a genre over there called ‘boyar-anime’ which is essentially fantasy set in imperial russia. I do most my heavy translation using Anthropic Haiku which is very cheap and unlike the higher end models it tends to dumb down some of the more complex parts of Imperial Russian aristocracy so it’s more in line with similar fiction over here. When I take the source book, I chunk it down into small segments that I translate individually so I don’t get context bleed, then I mechanically process to find anything that didn’t translate very well. I combine roughly 40 of these weirdly translated segments into a jsonl file and submit the file through the API. OpenAI Batch API can accept up to 900k tokens, but you’ll wait close to 11 hours for something that large. 40 segments is around 30k tokens and that usually processes in a few mins to an hour depending.
The jsonl file is essentially made up of smaller json blocks
{ "custom_id": "SEGMENT-NUM", "method": "POST", "url": "/v1/responses", "body": { "model": "gpt-5.3", "input": [ { "role": "system", "content": [ { "type": "input_text", "text": "You are a meticulous English language proofreader." } ] }, { "role": "user", "content": [ { "type": "input_text", "text": "PROMPT - SUBMITTED SEGMENT" } ] } ], "max_output_tokens": 8192 } }I then setup polling to check back with the API every few mins, when the submitted queries are completed, I send more automatically until everything has been processed.