Comment by unicornfinder

The Dumb Zone for Opus has always started at 80-100k tokens. The 1M token window just made the dumb zone bigger. Probably fine if the work isn't complicated but really I never want an Opus session to go much beyond 100k.

Claude Opus can use a 1M context window but I’ve found it to degrade significantly past 250k in practice.

The cost per message increases with context while quality decreases so it’s still generally good to practice strategic context engineering. Even with cross-repo changes on enterprise systems, it’s uncommon to need more than 100k (unless I’m using playwright mcp for testing).

I had thought this, but my experience initially was that performance degradation began getting noticeable not long after crossing the old 250k barrier.

So, it has been convenient to not have hard stops / allow for extra but I still try to /clear at an actual 25% of the 1M anyhow.

This is in contrast to my use of the 1M opus model this past fall over the API, which seemed to perform more steadily.

I’m genuinely surprised. I use copilot at work which is capped at 128K regardless of model and it’s a monorepo. Admittedly I know our code base really well so I can point towards different things quickly directly but I don’t think I ever needed compacting more than a handful in the past year. Let alone 1M tokens.

The context windows of these Chinese open-source subscriptions (GLM, Minimax, Kimi) is too small and I'm guessing it's because they are trying to keep them cheap to run. Fine for openclaw, not so much for coding.

Personal opinions follow:

Claude Opus at 150K context starts getting dumber and dumber.

Claude Opus at 200K+ is mentally retarded. Abandon hope and start wrapping up the session.

Don’t want to disappoint you, but above 200k opus memory is like a gold fish. You need to be below 150k to get good research and implementation.

Quality degrades fast with context length for all models.

If you want quality you still have to compact or start new contextes often.

I haven't screenshotted to alas, but it goes from being a perfectly reasonable chatty LLM, to suddenly spewing words and nonsense characters around this threshold, at least for me as a z.ai pro (mid tier) user.

For around a month the limit seemed to be a little over 60k! I was despondent!!

What's worse is that when it launched it was stable across the context window. My (wild) guess is that the model is stable but z.ai is doing something wonky with infrastructure, that they are trying to move from one context window to another or have some kv cache issues or some such, and it doesn't really work. If you fork or cancel in OpenCode there's a chance you see the issue much earlier, which feels like some other kind of hint about kv caching, maybe it not porting well between different shaped systems.

More maliciously minded, this artificial limit also gives them an artificial way to dial in system load. Just not delivering the context window the model has reduces the work of what they have to host?

But to the question: yes compaction is absolutely required. The ai can't even speak it's just a jumbled stream of words and punctuation once this hits. Is manual compaction required? One could find a way to build this into the harness, so no, it's a limitation of our tooling that our tooling doesn't work around the stated context window being (effectively) a lie.

I'd really like to see this improved! At least it's not 60-65k anymore; those were soul crushing weeks, where I felt like my treasured celebrated joyful z.ai plan was now near worthless.

There's a thread https://news.ycombinator.com/item?id=47678279 , and I have more extensive history / comments on what I've seen there.

The question is: will this reproduce on other hosts, now that glm-5.1 is released? I expect the issue is going to be z.ai specific, given what I've seen (200k works -> 60k -> 100k context windows working on glm-5.1).