Max_output_tokens when using gpt 5.5 xhigh detailed reasoning

API
1

I am using chatgpt 5.5 xhigh and I am hitting an error I hope i can get some clarification on; it say max_output_tokens when reasoning;

It says I sent in around 271,000 tokens, there’s around 89 tokens, it starts to reason and stops;

Can I continue from this phase?

here’s what the server is sending back over sse right around when it decides to just stop

"usage": {
  "input_tokens": 271467,
  "output_tokens": 79,
  "output_tokens_details": {
    "reasoning_tokens": 79
  },
  "total_tokens": 271546
}
2

Hi,

and welcome to the developer community.

max output token means you have reached the limit of the model.

You should split the content to reason over it preferably building a knowledgegraph with an analyzer pipeline.

Or try to split the content, ask a model to summarize it and extract key knowledge..

But without knowing what it is it is hard to tell.

Or in short to answer your question: nope.

3

Thanks for the feedback, I looked at the codex github repo and they use websockets and pass a previous response id, would that get around this with a websocket connection or is it a hard limit?

I don’t want to go down that route only to hit the same limit, is this just something that’s a system limit? I build a stateless ledgar that I send; I can drop parts or compact it but it seems to always fail around this 270,000 token with between 70-120 reasoning token; the complexity to implement stateful compaction is a bit high, I just want to clarify before I decide how to proceed.

4

Are you doing this on codex or directly through the responses API?

I think codex has a forced lower limit to avoid usage to drain too fast.

5

I am doing it through the api; not directly through the codex tui/ desktop app.

6

Large input context models are gpt-5.4+, with 1050000 context window possible.

GPT-5.x models prior are a 400,000 context window, of which 272k can be sent as input, with the 125k remaining reserved as output context generation.

Sending more than 272k tokens to a large context model will trigger double-the-price, somewhat “turning on” long context mode.

What you show and experience should not be the case, but instead would be a bug: the model’s “short” total context window is not a maximum 272000, but instead 400k as before. All gpt-5 models state a 128k output, from which the input shall not consume, but should be reserved for generating an output for you.

“Long context” vs you running “Short context” on an input under the high-price switch shall not be an off switch of output of only 272k total context window if you do not trigger it with a large input. It shall perform like gpt-5.2 across all models, always providing you output space.

What you could try: sending the API parameter max_output_tokens, and setting it much higher than you might expect. This is the total budget for an API call’s output, which not only includes reasoning tokens, but also when internal tools are used, collects all generated tokens into a final report.

Then if running up against a bug with new models gpt-5.5 or gpt-5.6, you could complete the $1.36 call there against gpt-5.2, a cheaper input price that has no alternate “long” mode that might have been provisioned incorrectly.

Also: watch and collect the entire API response object: any refusal or error output items, status “completed” or not, or the report in the field “incomplete_details”. That may give a bigger picture of what is terminating the output before a seen response.