Latest / AI Fire Daily / #506 Neil: GLM-5.2 Closes The Gap With Claude And GPT For Way Less Money
Transcript
- 0:00A developer recently gave a new AI model a coding
- 0:03task. And instead of instantly spitting out a
- 0:07wall of text, the model did something eerie.
- 0:12It paused. It mapped out the architecture deeply.
- 0:15It actually asked a few clarifying questions
- 0:17first. Right, which is wild. And it built a clear
- 0:20plan of attack. Right. It behaved exactly like
- 0:23a senior software engineer. It really is a fundamental
- 0:25shift in machine behavior. You know, you are
- 0:27watching basic pattern matching evolve into true
- 0:30reasoning. Welcome to the Deep Dive. We are exploring
- 0:33something truly fascinating today. Absolutely.
- 0:36So Anthropic recently pulled Claude Fable 5 from
- 0:39public access. And this was due to a new export
- 0:42control order. Yeah, things are getting highly
- 0:44restricted out there. Right. But ZAI did the
- 0:47exact opposite. They just released their new
- 0:49model, GLM 5 .2. Completely out of left field,
- 0:52too. And it is completely open. It is remarkably
- 0:54cheap, and it is entirely free to try right now.
- 0:58It really is a profound contrast in industry
- 1:00direction. You know, one massive door closes
- 1:03tightly and another door kicks wide open. Today,
- 1:05we will unpack this new Frontier model. We are
- 1:08going to explore its shocking benchmark wins.
- 1:12We will break down its highly disruptive pricing,
- 1:15and we will explain exactly how you can use it
- 1:18today. It's a critical moment for open weight
- 1:20models. You really need to understand what is
- 1:22happening here. OK, so let us start with the
- 1:24underlying machinery. What makes this senior
- 1:27engineer behavior actually possible? Right. ZAI
- 1:30built GLM 5 .2 as their absolute flagship model.
- 1:35It focuses heavily on three main areas. Yeah,
- 1:37advanced coding, complex agent tasks, and long
- 1:41horizon work. That last point is incredibly important
- 1:44right now. Oh, it's everything. I mean, many
- 1:46models excel at giving quick short answers, you
- 1:49know, trivia or whatever. Right. But they completely
- 1:51fail during a long complex workflow. GLM 5 .2
- 1:55is designed to maintain strict reliability. It
- 1:57stays focused through days of sequential complex
- 2:00steps. Exactly. It doesn't get lost in the weeds.
- 2:03So it relies on three core pillars of architecture.
- 2:06The first is a 1 million topin context window.
- 2:09Which is huge. That is an enormous amount of
- 2:11raw information. It really is. Yeah. I mean,
- 2:13you can load a massive code base inside it. You
- 2:16process the entire project in one single pass.
- 2:19Wow. Yeah, you do not have to chop files up anymore.
- 2:21It just reads the whole thing. The second pillar.
- 2:23is an open MIT license. Super permissive. Right.
- 2:27And the third pillar involves its two distinct
- 2:30reasoning modes. They are called high and max.
- 2:33We will definitely explore those specific modes
- 2:36later on. But first, let us look at its actual
- 2:38size. It features 753 billion total parameters.
- 2:44Massive. But, and this is the interesting part,
- 2:46it uses a mixture of experts set up. Right, the
- 2:50MoE architecture. Yeah, so it only activates
- 2:5240 billion parameters per token. Beat. Can you
- 2:56explain what a mixture of experts set up actually
- 2:58is? Yeah, simply put, it only activates the parts
- 3:01of its brain needed for the task. OK, so it operates
- 3:04kind of like a massive university campus. I love
- 3:07that analogy, yeah. You do not? consult the entire
- 3:09school for one single problem. You just walk
- 3:12into the specific department you need. Exactly.
- 3:14You visit the physics lab for a gravity problem.
- 3:16Right. You ignore the English department entirely.
- 3:19You don't need them for that, so you save the
- 3:21energy. But wait, I have to push back on this
- 3:23a little bit. Sure. If it is a 753 billion parameter
- 3:27model, but only using 40 billion at a time, are
- 3:31you losing the deep intelligence of the other
- 3:33700 billion? How does it actually know it is
- 3:36picking the right experts? Well, that is the
- 3:39genius of the internal routing network. The model
- 3:42has this specialized gateway layer. It analyzes
- 3:46your prompt. instantly before processing the
- 3:49actual answer. It mathematically calculates which
- 3:52specific expert clusters hold the relevant knowledge.
- 3:55Oh, wow. So you are not losing the intelligence
- 3:57at all. You are just filtering out the noise.
- 3:59You keep the raw strength of a massive system,
- 4:02but you run it much faster overall. Exactly.
- 4:05And you use significantly fewer computational
- 4:07resources doing it. It is highly efficient. Highly,
- 4:10yeah. This leads me to a really critical question.
- 4:11OK. Why does an open MIT license matter so much
- 4:14right now? API access is incredibly ubiquitous
- 4:18across the tech industry. You can just rent access
- 4:21anywhere. Well, because you are completely independent
- 4:24from centralized corporate gatekeepers. Right.
- 4:27An MIT license offers a very permissive legal
- 4:29framework. You get the absolute freedom to host
- 4:32the model yourself. You can edit the underlying
- 4:34weights directly. Exactly. You never have to
- 4:37worry about a company pulling access overnight,
- 4:39like we just saw with Anthropic. Ah. So you own
- 4:42the engine entirely instead of just renting the
- 4:44steering wheel. That's a perfect way to put it.
- 4:46Yeah. Right. And it guarantees your long -term
- 4:48stability. You can actually build a business
- 4:50on it safely. The theory makes perfect sense.
- 4:52Yeah. But a massive architecture is only useful
- 4:57if it actually delivers results. True. We need
- 4:59to see how this lightweight expert performs.
- 5:02Let us compare it. against the major industry
- 5:05giants. Let's do it. So the baseline comparison
- 5:08is its own predecessor, GLM 5 .1. Right. The
- 5:11new benchmarks show a massive jump forward. It
- 5:14is not just a minor incremental update here.
- 5:17It is a paradigm shift. It shows clear gains
- 5:20in three distinct areas. Coding, advanced tool
- 5:23use, and deep reasoning. Yeah, so on Terminal
- 5:25Bench 2 .1, the old model scored 62 .0. The new
- 5:30GLM 5 .2 scores a solid 81 .0 points. That is
- 5:35a very respectable leap in performance. It is.
- 5:37But the Deep SWE score is truly shocking to me.
- 5:40Oh, yeah. That one is wild. GLM 5 .1 only hit
- 5:4418 points, though, on this benchmark. Which was,
- 5:47you know, fine at the time. Right. But GLM 5
- 5:50.2 jumps all the way to 46 .2. It more than doubled
- 5:53its previous capability score. That is incredible.
- 5:56It is. And keep in mind, DeepSWE measures how
- 5:59a model handles much harder tasks. It is not
- 6:02just writing simple Python scripts from scratch.
- 6:05It is fixing deep, complex architectural bugs
- 6:08within existing repositories. Real -world engineering.
- 6:10Exactly. Now let us compare it to the closed
- 6:12models. On the Long Horizon Frontier SWE benchmark,
- 6:16GLM 5 .2 hits 74 .4%. Which is deeply impressive
- 6:20for an open model. It actually passes GPT 5 .5,
- 6:23which sits at 72 .6%. And it trails Claude Opus
- 6:264 .8 by less than a single point. Right. Opus
- 6:29is currently at 75 .1%. It is a highly nuanced
- 6:32picture, though. Oh, for sure. It does not win
- 6:34every single category across the board. No, it
- 6:36doesn't. Like on Post Train Bench, GLM 5 .2 scores
- 6:3934 .3%. Right. This is lower than Opus 4 .8 at
- 6:4337 .2%. But it remains higher than Opus 4 .7.
- 6:48True. It also beats GBT 5 .5 and Gemini 3 .1
- 6:51Pro there. And on SWU Marathon, GLM 5 .2 holds
- 6:55second place overall. It scored 13 .0 % on that
- 6:59specific evaluation. Yeah, it sits slightly behind
- 7:02Opus 4 .8 and Opus 4 .7. But it stays ahead of
- 7:05GPT 5 .5 and Gemini 3 .1 Pro. This proves it
- 7:08is a true frontier level competitor. It trades
- 7:12blows with the absolute best models available
- 7:14anywhere. It does. Then we arrive at the biggest
- 7:17surprise of all. Oh man, the design test. Yeah.
- 7:19The model was tested on Design Arena. Right.
- 7:21This is a community -voted benchmark for web
- 7:24design. It focuses heavily on front -end generation
- 7:26and visual aesthetics. Which is usually a weak
- 7:29point. Yeah. But GLM 5 .2 achieved first place
- 7:32on this entire leaderboard. First place. Yeah.
- 7:34It earned a highly respected 1360 ELO score.
- 7:36Wow. It actually moved ahead of Claude Fabel
- 7:395. Whoa. Imagine an open -weight model not just
- 7:43writing backend code, but generating interfaces
- 7:46with actual visual appeal. It's crazy. It is
- 7:49usually all logic and absolutely no style. Yeah,
- 7:52it is a massive breakthrough for open source
- 7:54AI. Open models usually struggle deeply with
- 7:57visual aesthetics. Why is that? Well, there is
- 8:00a clear reason for that. They are trained heavily
- 8:03on scraping raw code repositories like GitHub.
- 8:06Right, right. So they learn pure functionality,
- 8:09but they do not learn human -tested visual layouts.
- 8:12They can write the underlying functional logic
- 8:14perfectly. But the buttons and layouts usually
- 8:17look terrible. Exactly, just totally unstyled
- 8:20blocks. So how does an AI model actually get
- 8:22graded on something as subjective as visual appeal?
- 8:25Ah, so the ELO system relies entirely on crowdsourced
- 8:29human voting. Really? Yeah. Real developers look
- 8:31at two different front -end designs blindly.
- 8:34They vote on which interface actually looks better
- 8:36visually. It bypasses automated, rigid grading
- 8:39completely. So humans blindly vote on layouts,
- 8:42teaching the model actual subjective aesthetic
- 8:44taste. Right. It relies on genuine human preference.
- 8:48That is fascinating. Yeah. Beating top models
- 8:51in benchmarks is one thing entirely. Doing it
- 8:53without raising prices is something else completely.
- 8:56That is where the industry really feels the economic
- 8:59impact. It goes against the current market trend
- 9:01completely, doesn't it? Oh, 100 percent. Most
- 9:03AI companies raise prices with every new capability.
- 9:065 .2 kept the exact same API pricing. It's wild.
- 9:11It costs the same as the older GLM 5 .1. Let
- 9:15us break down the exact economic numbers here.
- 9:17Okay, so input tokens cost $1 .40 for every 1
- 9:20million. Output tokens cost $4 .40 for every
- 9:241 million. And cached input is exceptionally
- 9:27cheap to use. It costs just $0 .26 for every
- 9:301 million tokens. Right. They also have a limited
- 9:32time free offer on cash storage. Let us compare
- 9:34this directly to the competition. Sure. Claude
- 9:36Opus 4 .8 charges $5 for input tokens. OK. And
- 9:40it charges a massive $25 for output tokens. Wow.
- 9:44GPT 5 .5 also charges $5 for input tokens. Yeah.
- 9:48But it charges $30 for output tokens. That is
- 9:51a huge jump. And Claude Fable five used to charge
- 9:54even more. It was $10 for input and $50 for output.
- 9:58That was before it was pulled from public access.
- 10:00The price gap here is absolutely staggering to
- 10:03see. It really is. I mean, think about it. If
- 10:05you are a startup. building an AI agent, making
- 10:09thousands of calls, a $30 output versus a $4
- 10:12output completely changes your business model.
- 10:15It is the difference between a viable product
- 10:17and bankruptcy. Exactly. It places GLM 5 .2 safely
- 10:21in the mid -range tier economically. But it delivers
- 10:23top tier, front tier level reasoning capabilities.
- 10:26Z .AI is clearly prioritizing long -term market
- 10:29adoption here. They want developers embedding
- 10:32this deeply into new products. They also launched
- 10:34a specific GLM coding plan. Right, the subscription.
- 10:37Yeah, it is a separate subscription built specifically
- 10:39for software development. It starts at $18 a
- 10:41month for the light tier. And they have pro and
- 10:43max tiers for heavier workloads. It is tailored
- 10:46for high frequency complex coding tasks. Exactly.
- 10:50But with prices this low, is the API meant for
- 10:53enterprise scaling or just personal testing?
- 10:56The API is explicitly built for heavy product
- 10:59integration. Oh, really? Yeah. It allows developers
- 11:01to scale massive automation workflows rapidly.
- 11:04You get frontier performance without destroying
- 11:07your monthly operational budget. Got it. The
- 11:10chat is a playground, but the API drives scalable
- 11:13business engines. Exactly. It changes the fundamental
- 11:16math for startups to sex silence. Sponsored.
- 11:20Now that we understand the raw power and price,
- 11:23let us talk about how you can actually use it.
- 11:25Because you can get your hands on it today, easily.
- 11:27You have a few different access options available,
- 11:29but first let us detail those two reasoning modes.
- 11:32Right. You mentioned high and max earlier in
- 11:34the show. Yeah, so GLM 5 .2 lets you pick how
- 11:36deep it actually thinks. OK. You control the
- 11:39speed, cost, and quality all at once. The high
- 11:42mode is fast and perfectly balanced. You turn
- 11:44this on through the thinking parameter settings.
- 11:47Exactly. It answers much faster, while maintaining
- 11:50very good quality. It is perfect for everyday
- 11:53tasks like writing emails or generating simple,
- 11:58straightforward code snippets quickly. Then you
- 12:00have the Max Reasoning mode available. Yeah,
- 12:03the heavy little term. Max pushes the model to
- 12:05think much, much deeper. It pushes the extreme
- 12:08limits of logical problem solving. You use this
- 12:11for complex, multi -file code refactors. Right.
- 12:15Or incredibly hard debugging sessions inside
- 12:17messy legacy code bases. So you trade some speed
- 12:21for noticeably sharper, accurate answers. Right.
- 12:24You turn it on using the reasoning effort parameter.
- 12:27And you can also add web search. to these requests.
- 12:31Yes, which is huge. It costs just one single
- 12:34penny per search call. Pairing web search with
- 12:36MaxMode is incredibly powerful. You get deep
- 12:39thinking combined with fresh current information.
- 12:42There are three main ways to access this model.
- 12:44Let's rank through them. Option one is the official
- 12:47z .ai chat interface. Right. It is completely
- 12:49free to use right now. You do not even need an
- 12:51API key. And it works seamlessly right inside
- 12:53your web browser. You just pick GLM 5 .2 from
- 12:56the dropdown model list. It even supports large
- 12:58file uploads up to 10 megabytes. Which is great
- 13:01for documents or small code bases. This is best
- 13:03for someone testing the waters first. Definitely.
- 13:06Option two is utilizing the Alamacloud platform.
- 13:09Yeah, this specific route... does require a paid
- 13:12monthly subscription. But it runs the model on
- 13:14U .S. hosted GPU infrastructure. Right. And it
- 13:17is highly focused on strict data privacy protocols.
- 13:20Your prompts and responses do not get logged
- 13:22anywhere. That's vital for a lot of companies.
- 13:25Yeah, they are never used to train the hodl again.
- 13:27One important note about this specific Elama
- 13:30option. The model only runs through their cloud
- 13:33infrastructure servers. You cannot self -host
- 13:36it on your own normal hardware. Right. The model
- 13:38size goes well beyond what normal computers handle.
- 13:41Then option three is using the open router platform
- 13:44ecosystem. This is where you create your own
- 13:46API key. You call GLM 5 .2 through a standard
- 13:50access endpoint. The system routes your request
- 13:53to the best server. It optimizes automatically
- 13:56based on speed or current price. It works perfectly
- 13:59if you use several different models. You manage
- 14:01all of them through one single access point.
- 14:04So real world application demos are where things
- 14:07get really exciting. Oh, absolutely. Benchmark
- 14:09numbers only tell you part of the true story.
- 14:11We need to see what happens in actual practice.
- 14:14And the community has been testing it hard. Demo
- 14:171 was a side -by -side test on X. Right. They
- 14:19used the exact same prompt and reference image.
- 14:22They asked GLM 5 .2 and Cloud Opus 4 .8 to compete.
- 14:26And the task was building a 3D logistics dashboard
- 14:29interface. Both models successfully produced
- 14:32a working 3D interactive globe. Which is hard
- 14:35enough. Right. They used the 3JS library to build
- 14:39the complex visualization. Yeah, and it featured
- 14:41complex route arcs and transparent glass panels.
- 14:44It looked incredible. The final layout was highly
- 14:46polished and professional looking. GLM 5 .2 held
- 14:49its own completely against Opus 4 .8. Which is
- 14:52crazy because complex 3D interface work is usually
- 14:55dominated by closed models. Exactly. But Demo
- 14:582 is even more fa - fascinating to look at. OK,
- 15:00tell me about that one. Someone ran the model
- 15:02fully local on a Mac Studio. Fully local. Yeah,
- 15:05they used a heavily compressed two bit quantized
- 15:08version. Let us pause here for a second. How
- 15:11exactly does quantization shrink a massive model
- 15:15down? Think of it like compressing a high resolution
- 15:19raw photograph. You shrink it down into a smaller
- 15:21JPEG file. Quantization strips out the extreme
- 15:25decimal precision of the model's weights. You
- 15:29lose a tiny fraction of nuance, but it reduces
- 15:32the memory footprint drastically. Oh, I see.
- 15:35This allowed it to fit on personal consumer hardware.
- 15:38Right. And the real world results actually came
- 15:40out better than Opus 4 .8. Wow. It now powers
- 15:43a Hermes agent and codec set up locally. It is
- 15:46completely free and totally private on that machine.
- 15:49I still wrestle with setting up local quantized
- 15:51models on my own machine. Oh, absolutely. I mean,
- 15:54the dependency errors alone will drive you crazy
- 15:55sometimes. It is definitely a highly technical,
- 15:58frustrating process, but the ultimate payoff
- 16:01is having complete, uncompromised privacy. Right.
- 16:04But if the free chat does so much, when does
- 16:07a user actually hit the wall and need the Olamacloud
- 16:10or OpenRouter options? You hit that wall when
- 16:13you need deep automation. You need APIs to integrate
- 16:16the model into custom software, or when your
- 16:19personal hardware simply cannot run quantized
- 16:22versions. You upgrade when you need automated
- 16:24workflows or your laptop just can't handle it.
- 16:26Exactly. The chat is great, but APIs build actual
- 16:30products. Let us zoom out and recap the big ideas
- 16:32here. Sounds good. What is the core narrative
- 16:35of this entire deep dive? GLM. 5 .2 is much more
- 16:40than an open source alternative. It is a massive
- 16:43753 billion parameter mixture of experts frontier
- 16:48model. It achieves top tier reasoning capabilities
- 16:51across the board. It designs beautiful UI prototypes
- 16:54that look truly finished. And it does this at
- 16:57a fraction of the cost. It keeps the entire AI
- 16:59ecosystem highly accessible today. Especially
- 17:02while gatekeepers are actively tightening their
- 17:04grips on advanced models. ZAI is providing a
- 17:07powerful, stable, open alternative right now.
- 17:10It does not fully replace closed models in every
- 17:12single scenario. No, not every single one. But
- 17:14with its price and openness, it earns a spot.
- 17:16It belongs in every serious developer's modern
- 17:19toolkit today. It is a fascinating evolution
- 17:21of the open source landscape. It really is. And
- 17:24it leaves you with a very profound final thought.
- 17:26OK. A free open weight model can now pause and
- 17:29reflect. Yeah. It can ask deeply clarifying questions
- 17:32about a prompt. It can map out a complex software
- 17:35architecture project. It does all this before
- 17:38writing a single line of code. It's unbelievable.
- 17:42Essentially, it is flawlessly mimicking a senior
- 17:44human engineer. So what exactly does the role
- 17:49of a junior human developer look like five years
- 17:51from now? Man, that is the defining technical
- 17:54question of our generation. Thank you for joining
- 17:57us on this Deep Dive. Keep questioning the tools.
- 17:59Keep exploring the frontier. Out to your own
- 18:01music.