This is the same underlying limitation as your Kimi K3 thread, and I’ve added this report to the issue we’re tracking. Gemini 3.8 Flash currently has no Context option in the model picker, so every chat runs with the standard 200K working window (unless it has gotten stuck in the now-invisible Max Mode, which bumps it up to 1M).
In that case, both Gemini and Kimi become useless for long tasks. For example, right now I consider Kimi to be the ideal model for a pipeline where I might have 300,000–400,000 tokens at the end, and those tokens can’t be compressed because the pipeline breaks down.
OpenAI’s 1M models are slightly worse for these tasks than K3, and on top of that, they may become unavailable in November.
