Comment by RickHull
5 hours ago
I am on their "Coding Lite" plan, which I got a lot of use out of for a few months, but it has been seriously gimped now. Obvious quantization issues, going in circles, flipping from X to !X, injecting chinese characters. It is useless now for any serious coding work.
I'm on their pro plan and I respectfully disagree - it's genuinely excellent with GLM 5.1 so long as you remember to /compact once it hits around 100k tokens. At that point it's pretty much broken and entirely unusable, but if you keep context under about 100k it's genuinely on par with Opus for me, and in some ways it's arguably better.
100k tokens it's basically nothing these days. Claude Opus 4.6M with 1M context windows is just a different ball game
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.
1 reply →
Quality degrades fast with context length for all models.
If you want quality you still have to compact or start new contextes often.
Is manual compation absolutely mandatory ?
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).
2 replies →
I am on the mid tier Coding plan to trying it out for the sake of curiosity.
During off peak hour a simple 3 line CSS change took over 50 minutes and it routinely times out mid-tool and leaves dangling XML and tool uses everywhere, overwriting files badly or patching duplicate lines into files
Off peak for China or US
Off peak for China. Off peak times are only in one timezone
My impression is that different users get vastly different service, possibly based on location. I live in Western Europe, and it works perfectly for me. Never had a single timeout or noticeable quality degradation. My brother lives in East Asia, and it's unusable for him. Some days, it just literally does not work, no API calls are successful. Other days, it's slow or seems dumber than it should be.
Every model seems that way, going back to even GPT 3 and 4, the company comes out with a very impressive model that then regresses over a few months as the company tries to rein in inference costs through quantization and other methods.
This is surprising to me. Maybe because I'm on Pro, and not Lite. I signed up last week and managed to get a ton of good work done with 5.1. I think I did run into the odd quantization quirk, but overall: $30 well spent
I'm also on the lite plan and have been using 5.1 for a few days now. It works fine for me.
But it's all casual side projects.
Edit: I often to /compact at around 100 000 token or switch to a new session. Maybe that is why.
I'm on their lite plan as well and I've been using it for my OpenClaw. It had some issues but it also one-shotted a very impressive dashboard for my Twitter bookmarks.
For the price this is a pretty damn impressive model.
Is there any advantage to their fixed payment plans at all vs just using this model via 3rd party providers via openrouter, given how relatively cheap they tend to be on a per-token basis?
Providers like DeepInfra are already giving access to 5.1 https://deepinfra.com/zai-org/GLM-5.1
$1.40 in $4.40 out $0.26 cached
/ 1M tokens
That's more expensive than other models, but not terrible, and will go down over time, and is far far cheaper than Opus or Sonnet or GPT.
I haven't had any bad luck with DeepInfra in particular with quantization or rate limiting. But I've only heard bad things about people who used z.ai directly.
I'm on their Lite plan and I see some of this too. It is also slow. I use it as a backup.
> Obvious quantization issues
Devil's advocate: why shouldn't they do it if OpenAI, Anthropic and Google get away with playing this game?
I think what Anthropic is doing is more subtle. It's less about quantizing and more about depth of thinking. They control it on their end and they're dynamically fiddling with those knobs.
It has been useless for long time when compared to Opus or even something like Kimi. The saving grace was that it was dirt cheap but that doesn't matter if it can't do what I want even after many repeated tries and trying to push it to a correct solution.