Comment by Tiberium
13 hours ago
It seems to really be a nice step-up and is getting quite close to the frontier. I wish they'd start focusing on the reasoning efficiency now, though. I have a simple (relatively) test task to evaluate LLMs: writing a simple math evaluator library in Nim (it's about 400-600 lines total max), and GLM 5.2 (xhigh which maps to max effort) spent over 15 minutes (!) reasoning, spending about 45k tokens, before it finally wrote the first file.
I know it's hard to improve on that, but now that their models are good enough at raw intelligence, I think this should become a higher priority task.
Currently on https://artificialanalysis.ai/#output-tokens GPT 5.5 xhigh spends 16k tokens total on average, GPT 5.5 high is 10k, Fable 5 33k, Opus 4.8 41k, GLM 5.2 is 42k. GPT 5.5 is extremely reasoning efficient.
Of course if you convert those values to actual request cost, GLM 5.2 will probably beat GPT 5.5/Opus 4.8, but speed matters for a lot of people, I think.
GLM 5.2 Max = Opus 4.8 Max in thinking behavior. The thinking chain is so similar, and so is the amount of token usage on the output.
If you want reasonable token usage, you need to run it GLM 5.2 at High. There is little drop in quality from Max to High (for most tasks). And it cuts token usage by 2 a 2.5x. GLM 5.2, Max is really something you only need for complex tasks.
In essence, GLM 5.2 is Opus 4.8 its little brother, at a way, WAY cheaper price.
There has been really no training on Opus models going on, really, none i tell you! /sarcasm
> GLM 5.2 Max = Opus 4.8 Max in thinking behavior
This is insane! I can't wait until technology progresses to the point we can run these things on consumer hardware!
you need 8 x 96GB Blackwell or equivalent
so around US$150k which is Small/Medium-Enterprise territory already, but who knows when it will hit "reasonable" home consumer territory
I think there's hope future generations of unified memory machines may get this sort of memory availability when new fabs open in then next couple of years and then ramp up production for a few years afterwards - that makes ~2030s credible at this point, but nobody can really predict the market that far ahead
1 reply →
Are there any indications that this will be possible? Consumer hardware will continue getting better but I can't see 512GB RAM in a MacBook Pro any time soon. I'm hoping linear attention techniques plus MoE will make breakthroughs in size/compression and throughput.
2 replies →
distillation of thinking models is not particularly effective - both "Open"AI and Misanthropic don't show you the real chain of thought, only its severely downscaled version. both do everything in their power to combat such outrageous copyright infringement, so the bulk of unethically scrapped data the Chinese have is from several generations ago.
It is quite likely that the intermediate tokens don’t have ‘semantic import’[0]
There are methods like Habitual Reasoning Distillation or Inverted Reasoning Traces [1] that can help.
While there are reasons to hide the intermediate tokens from a IP protection stand point, there is also a need to hide more effective and efficient generating that doesn’t fit the R1 claims of an aha moment that has been debunked, but is a consumer expectation.
While hidden intermediate tokens do increase the difficulty, it is not a from barrier in itself, especially as they are billed, given information about their length.
[0] https://arxiv.org/abs/2504.09762v4
[1] https://arxiv.org/abs/2603.07267
>such outrageous copyright infringement
Sarcasm, considering the source of their own training data?
3 replies →
For Claude models at least, you can tell to just manually think in the output and it works fine. I do it reguralrly because for creative writing and summarization, they seem to believe they don't need to think at all, and get way worse results.
1 reply →
FYI: model outputs are not protected by copyright.
Chinese distillation attacks are about as unethical as Robin Hood stealing from the rich to give to the poor. The real unethical scraping was done by Anthropic to train Claude.
To be clear, if Anthropic was using totally licensed data, I'd be sympathetic to these claims. But if you're going to pirate the world's creativity you'd better be willing to gimme dat shit for free[0].
[0] As said by Hungry Santa.
I don’t understand why there isn’t public dataset for reasoning that can be improved by humans/llms like Wikipedia (ie with auto judging contributions etc).
1 reply →
Reasoning models can coaxed to reason like they do in dedicated reasoning blocks, outside of those blocks: in normal parts of the response.
But Anthropic at least has openly admitted they try to detect that and interfere
Supposedly there are “jailbreaks” that expose considerably more of the thinking traces.
You can trivially leak the CoT of any current model, it's not a problem.
>outrageous copyright infringement
>unethically scrapped data
Hahahahaha
The companies that did copyright infringement and unethically scrapped data think that copyright infringement and unethically scrapping data is wrong and needs to be stopped.
Though only in particular situations, like when it’s done to them and not when they do it. Cause they have the power and are morally right and know better than you. And if you question this at all, well you’re a threat to American values and a supporter of the Chinese and leading to the break down of Democracy.
This isn’t a type of reasoning argument or manipulation tactic used by the rich throughout history to trick the naive and gullible masses or anything like that. Trust me, I’m rich and I’m morally right. /sarcasm
looking at the score this is rather a gemini 3.5 flash competitor, yes, for cheaper, but distance to opus and fable is as big as their price diff.
With such ridiculously long thinking traces I'm surprised max outperforms high. After all, performance falls off a hill after a certain amount of context, and long thinking traces can fill that up really quickly.
> It seems to really be a nice step-up and is getting quite close to the frontier.
IMHO it's already surpassed them. I vastly prefer my personal GLM and OpenCode setup to the Claude Code and Opus one that I have to use at work. The former makes way fewer StackOverflow brogrammer-tier mistakes and is considerably better at following instructions. The harness UX is also vastly superior as it doesn't ignore, randomly change, or incorrectly report settings.
Maybe it's the harness and I'd have even greater success with OpenCode and Anthropic, but I think it safe to say that Anthropic's moat is evaporating.
This is a problem I find with opus is will spend so long thinking then going “but wait what if”
To point where I stop it and simple tell it to “start writing code you can work it out as you go along”
Seems writers block also effects LLM
https://arxiv.org/abs/2606.00206
In this paper they nerf an LLMs ability to emit waffling thinking tokens like "wait", "but", "alternatively", and the models (they're old, small models in the paper) terminate reasoning faster and perform better. I bet Anthropic is tuning this on their backend.
I imagine Anthropic would rather train a small control model instead of resorting to sampling hacks
This is super cool. Do you know if any of the inference backends (llama.cpp, vllm, etc) support this technique?
1 reply →
I usually have Claude build a plan first, then I put it into an XML file it updates with phases, usually we talk about some of those tasks, and then once its good and I like it, I have Claude implement the plan.
Another thing I tell Claude to do is to not guess, but look at documentation, it messes up a lot less, might use some tokens reading docs, but at least it has a higher success rate code wise.
XML??
8 replies →
Seriously. Whenever I read the thinking output I get mad and turn down effort to medium or low.
Just output the code and we’ll work through it!
I feel similarly about having codex review claude’s plans. I don’t think I’ve ever seen it catch a major issue. It just points out things that would have inevitably been addressed during implementation anyway.
A lot of times this is how humans work. Just start 'putting words on paper', 'think by doing', etc. sometimes it's more efficient to see why something won't work after writing a bit of it, and sometimes you get lucky and it works right off the bat
Qwen is notorious for this, too. It’ll sometimes spin in a long loop of “But wait…” paragraphs.
Fable was 20 times worse on that.
It's clear it was the vibe coding model, as like no other model before, fully turned you into his assistant instead of the other way around.
Could it be possible, these firms are optimizing for two things: a) Better performance. b) Gathering data from you to further improve performance later. I've also found the huge amount of planning rather than iteration frustrating. I've felt like I'm teaching a junior!
5 replies →
I've been having success with Opus but you REALLY have to tame it. Long prompts that list what files to look at, relationships between entities, etc... I went from regularly hitting my daily limit to almost never hitting it. Oh, and also I was being lazy with small changes and stopping that helped a lot too. As you said, it gets in these loops where it's just churning and if you don't stop it it can go on for way too long.
Hopefully the recent work Moonshot did with Kimi K2.7 Code trickles in to the other open-model labs.
Per AA, while K2.7 Code is roughly on par w/ K2.6 in terms of intelligence, it uses half the output tokens to get there.
This is GLM 5.2 Max. GLM 5.2 High which use less than half[1] the tokens.
[1] https://z.ai/blog/glm-5.2
Yes, but the Artificial Analysis result is also from GLM 5.2 (max), not high.
They have this with a lot of models, measuring only the max setting, while the one you'd actually want to use for most tasks is much lower.
1 reply →
That's interesting. I gave nearly the same task to Gemma4 31b as a test yesterday. Write a symbolic math engine in Typescript that can perform evaluation and simple expression reductions over +-/*(). It performed the task correctly with minimal reasoning - much fewer reasoning tokens than output tokens.
Tbh, so what? I googled "symbolic math engine in Typescript that can perform evaluation and simple expression reductions over +-/*()" and got what looks to be viable answers without using any AI model at all. Reciting well established things from memory isn't terribly interesting. Show it a novel codebase and have it implement something within it.
TBH, while your point is a fair one, your attitude is off-putting and needlessly condescending.
So, a natural question would be why a model would ever get it wrong?
As per stats in other comments, it is frontier, not close to frontier.
> Of course if you convert those values to actual request cost, GLM 5.2 will probably beat GPT 5.5/Opus 4.8, but speed matters for a lot of people, I think.
GLM5.2 ends up being far more expensive than I thought it would be when I tried it on openrouter. I ground through $5 USD worth of tokens quite quickly.
And this was high, not max.
Using these open models really makes you realize how subsidized Anthropic and OpenAi's subscription plans are.
Absolutely. You can also run codeburn or ccusage and they'll scan the session files and tell you how much you burnt in API token pricing equivalent.
I agree. I've noticed that it is quite smart but it has a tendency to doubt itself and overthink. I monitor its internal dialogue and prod it when it does this. They need to optimize the chain of thought early stopping.