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
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.
> but I can't see 512GB RAM in a MacBook Pro any time soon
Could totally see this being a comment from a forum in like 1994 but swap out GB for MB and MacBook Pro to whatever the popular consumer pc was at the time
Well, we're probably not going to be running frontier models anytime soon, but I think the general assumption is smaller models will continue to improve until they're sufficiently good frontier models aren't needed.
There's potentially also augmentation through tools, harnesses and RAG to help boost how well they work without tons of parameters.
In the last ten years laptop memory footprints have, what, doubled at the low end? Smallest MacBook Pro in 2016 was 8GB, smallest is 16GB today? Max I think has gone up 8x meanwhile, 16 to 128?
I wonder if there's a bit of a chicken-and-egg issue where there wasn't much that demanded 10x the RAM, so there wasn't much pressure to develop more or increase production to support it at consumer prices.
There's wayyyyyyy more demand for memory generally now, so assuming it's not a demand bubble that pops rapidly, I'd expect the new normal to end up at a much higher baseline. 512GB would be 4x greater than today's max, so even with the relatively slow last 10 years development pace, give it five years max?
The problem is that the situation in the RAM market might just... not go away. It's locked in for the next couple of years unless the AI market goes pop. Which it might! But if it doesn't, there's no particular reason to think that the incentives for cornering the market like OpenAI have would go away.
We might see that new normal in five years or so. We will see a new normal sooner than that if there's a run on AI because of the sudden availability of DRR fab capacity, but also we'll probably see the level of local models freeze at whatever state they've got to at that point. But an equally likely outcome is that any new DDR capacity that comes online is just immediately absorbed by frontier AI, and consumer devices stay at "just good enough" for a decade.
The new Macbook Neo is 8GB. I think that if we are lucky, the huge RAM demand right now means new factory buildouts which eventually means more supply and prices go back down, and capacity begins to go up. This level of demand was just not anticipated by anyone.
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
> I think there's hope future generations of unified memory machines may get this sort of memory availability
I hope you're right. This is a very exciting idea. The weights are out there. The demand is astronomical. The manufacturers just need to make it happen.
there are cheaper ways to do it. not like, consumer-cheap, but I'm setting up a rig for 80% cheaper than that.
I'm a tad worried about triggering a run on the particular hardware I'm buying though so I'll leave it vague here, but hit me up on Discord if you're curious.
This is quite evident for personal AI but general intelligence with current scaling laws and how model keep getting better with more number of parameters, certainly the path does not converge.
Personal AI is more deprived of context today than quality of token. Having a on-system knowledge base paired with Gemma works well to large extend.
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.
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.
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].
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.
this helps so much. i do it too. with some of the newer frontier models its unclear if you can even turn it off in the first party chat apps. havent compared api semantics yet.
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
It’s been amazing to see the arc of tech people going from “evil Disney, copyright is an abomination, information wants to be free” to “OMG copyright is inviolable and AI is taking money out of Plato’s descendants’ pockets!”
> taking money out of Plato’s descendants’ pockets
Yeah, remind me - is it Plato's descendants that people are concerned about here, or is it every single author who had any work in Anna's Archive, any work published online, any work published on github, etc?
I think that people are probably upset about the harm to living people who had their work stolen by Meta and other LLM companies - regardless of license, terms of use, or any other attempted protection.
Sure, that’s the motte / bailey. Easy to point to living, starving writers who suffer grevious harm, in defense of perpetual copyright. Disney and others use literally this exact argument year after year.
I’m not even disagreeing. I’m just saying the shift in attitude about copyright in the tech space has been sudden, dramatic, and really funny. Remember “you wouldn’t steal a car”? Today’s anti-AI tech contingent are enthusiastically embracing that false equivalence that we all laughed at 20 years ago.
Having a static, immovable belief system about something like copyright that is unaffected by seismic shifts in the real world also doesn't seem very logical.
If like, Disney did a 180 overnight and bought rights from Google to scan every writer's saved work in Docs with some flimsy legal argument then a person saying "wait doesn't copyright actually protect that" would make sense. Even if you were previously upset about them suing schools for using 80 year art.
Simple trick: Use an agentic tool like Pi or OpenCode that allows you to switch models. First do some chats with DeepSeek or GLM who shows full thinking traces, then switch to Claude or GPT and it's more likely to show full thinking traces.
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).
For reasoning a manually-curated dataset is too small; you need to be able to automatically generate vast volumes of synthetic reasoning data with provably correct answers. That's presumably why Claude and GPT are so good at using Lean (the theorem prover), because they get fed a bunch of synthetic, verifiably correct training data.
> 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.
You would be surprised at how much of an impact the harness has. I switched to Pi and chinese open source models, and models that _I know_ are less capable than sonnet outperform my sonnet + claude code stack at work.
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 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.
Apparently because of how Claude is trained, even the system level prompts go through as XML, it works better with XML "prompting" so I figured I could have it write plans in XML. I need to update my ticketing tool to output XML maybe by default.
Comments later in thread say markdown works just as fine and that it’s more important to organize your plan into sections.
Also just think about it, why would a model trained on the world’s corpus of text (that isnt formatted in xml) perform better with XML? It would be a better study if that post tested markdown, org, xml, json, etc. 10 times to see if their is a difference
One reason to use XML-like formatting is that it makes the beginning and end of sections explicit. This is less of an issue when the model is generating text but can still be helpful when using templated prompts.
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
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!
I think they simply optimize around E2E benchmarks, none of those benchmarks is designed as multi turn assistance to the user, but going from a prompt straight to the final solution.
Exactly. How can "we" develop and encourage benchmarks for multi-turn user assistance?
That is what I want.
I feel like the models and harnesses push much too hard against this workflow -- that they push you towards letting go and vibe coding, with only your discipline (and desire for a quality and maintainable product) holding it back.
I think they are optimizing for one-shot performance because that will drive usage. They can’t afford to look bad in the benchmarks. And if that means consuming an order of magnitude more tokens, well, that’s good for business, too.
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.
I've been doing some testing with GLM 5.2 on Fireworks and it looks like the "High" reasoning level uses fewer tokens than even K2.7 Code by a considerable margin (roughly half).
Don't have any evals indicating how it compares on upper-bound quality, but for a well-defined task it seems like GLM 5.2 on "High" is remarkably token efficient. Looking forward to seeing where it lands on the AA index.
> 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.
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.
Agreed that models should get better at working with rare programming languages like Nim! Using them tends to confuse agents a lot in general. We're working on a paper right now where we compare how token-efficient models are when trying to implement the exact same program in different programming languages, and that's one of the trends we're seeing.
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.
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.