Note that GPT-5.6-Sol is ~2x as token efficient as Opus/Fable, so a 258K max is equivalent to ~516K on Claude models. You also may just be fooling yourself, as context rot is still a problem[1][2] and there's evidence compaction can be equivalent or better than longer context windows for agentic tasks[3]. Obviously it would be best if the models could reason over 1M context like they do over 256K context, but they can't yet.

[1] https://arxiv.org/abs/2605.12366 [2] see F1 of GraphWalks 256K vs 1M in Opus 4.8 System Card: https://www-cdn.anthropic.com/0b4915911bb0d19eca5b5ee635c80f... [3] https://context-folding.github.io/

The fact there is no way to disable auto-compaction like every other coding harness makes me scream. It fires randomly at 10-20%, so really you now only have guaranteed context of 272k * 80%. So many times I've had it work on a large codebase, it is almost done with a task, it needs maybe a 2000 token response, but it drops below 20%, it churns and churns and churns and then I see the dreaded ``Context compacted`` and scream. You can't go back in the history to the state of the conversation before the compaction. So then it has to investigate the codebase again gets to 20% and then auto-compacts. So it then has to read more, and then auto-compacts again, and then I'm out of tokens.

Full agreement. I sincerely hope that reducing token is mostly a cost saving measure and not some roundabout way to get more token use. It still does not help. In my corp, bean counters rendered initially useful internal llm largely pointless as they severely limited context.

I swear there is a executive convention out there sharing worst practices.

Just use md files for working memory. No need for large context. LLMs get dumber as you put more in context, as it stretches their attention. Keeping context small is better for quality.