# Long Responses API runs are costing us more from context than the output itself — retained record

> Retained reddit record from reddit:OpenAI.

Canonical: https://highsignal.app/data/records/5b80dc3e47f1cbed

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:OpenAI)  published 2026-10-07 16:22 ingested 2026-10-08 02:38

[original source ↗](https://www.reddit.com/r/OpenAI/comments/1x00xec/long_responses_api_runs_are_costing_us_more_from/)

## retained text



```
I’ve been looking closer at the token usage from one of our longer-running workflows using the Responses API and a big portion of the tokens on the expensive runs aren’t coming from what the model generates. The first couple steps are pretty normal but the workflow uses a few tools and keeps going from there and as the run gets longer we’re carrying more of the previous context and tool results forward so by the later steps the input side has grown quite a bit even though the model might only generate a relatively short response each time.

What makes it annoying is that these aren’t necessarily failed runs cause tey usually finish and give us what we need but two successful runs of the same workflow can end up with pretty different token usage depending on how many steps/tools were needed to get there.
I’m considering trimming older tool outputs or summarizing parts of the context after they’ve been used but I don’t want to save tokens to introduce more retries because the model lost something it needed.
If you guys are running longer workflows through the Responses API how are you handling this like are you pruning context during the run or accepting that the longer successful runs are going to be significantly more expensive?
```



document fetched 2026-10-08 00:29 · canonical https://reddit.com/r/OpenAI/comments/1x00xec/long_responses_api_runs_are_costing_us_more_from
