Latest / AI Fire Daily / #54 Robin: Open Source Just Hijacked Claude Code - Why GLM 5.2 is Replacing Opus 4.8 for 80% of Dev Work
Transcript
- 0:00A fully finished 23 second video intro beat generated
- 0:04from a single slash goal prompt. Wow. There was
- 0:07no step by step babysitting involved at all.
- 0:10It took one hour and 15 minutes to complete.
- 0:12Yeah, that is wild. And it didn't even max out
- 0:14the one million context window. Right. And it
- 0:17used three hundred fifty seven thousand tokens
- 0:19to do it. This completely shifts how we have
- 0:22to look at open source capabilities today. Welcome
- 0:25to the deep dive. We are very glad you are here
- 0:27with us. Absolutely. Today, our mission is to
- 0:30unpack a really rigorous hands -on guide. We
- 0:34are testing the open source GLM 5 .2 model, and
- 0:38we're doing this directly inside Cloud Code.
- 0:40We are going to walk through some actual real
- 0:42world tests here. We pit GLM 5 .2 against the
- 0:45heavyweight OPUS 4 .8. Right. We will outline
- 0:48a fascinating 80 -20 workflow split for your
- 0:51daily tasks. We will also break down the surprisingly
- 0:53simple non -technical setup. Finally, we will
- 0:56reveal why this specific shift makes open source
- 0:58AI impossible to ignore. Beat. But to really
- 1:02understand why this matters, we have to look
- 1:03at the data. Yeah, what actually happened when
- 1:06these two models went head -to -head in the wild?
- 1:08That is the most important question. There were
- 1:10several highly practical tests run here. Right.
- 1:14We wanted to see if GLM 5 .2 could genuinely
- 1:17compete on real work. Or did it just feel good
- 1:20because it was the cheaper option? Well, the
- 1:22first test was a one -shot web design task. The
- 1:25exact same prompt was used for both models. They
- 1:29both had to build a fully functional landing
- 1:31page. GLM finished the job in 3 minutes and 59
- 1:34seconds. Wow. And Opus took 14 minutes and 59
- 1:38seconds to finish. That is an absolutely massive
- 1:40difference in time. And GLM was around five times
- 1:43cheaper to run. Yeah. But the really fascinating
- 1:45part to me is the actual quality. Exactly. These
- 1:48were not clunky early 2000s wireframes. Right.
- 1:52Both results were highly polished and visually
- 1:55impressive. The pages had movement built into
- 1:58the design. They had properly structured sections
- 2:00and very clear calls to action. This was not
- 2:02some flimsy budget result you would have to rewrite
- 2:04anyway. No. It was a completely viable starting
- 2:08point for a real project. Looking at them side
- 2:10by side, you would have a hard time justifying
- 2:13that massive price tag for Opus. When the final
- 2:15result is that close, the price starts to matter
- 2:18heavily. It really does. But we didn't stop at
- 2:21visual design. Next up was hard coding. So GLM
- 2:24crushes basic web layout, which is great. But
- 2:28visual spacing is one thing. Logical reasoning
- 2:30is another entirely. Absolutely. That naturally
- 2:33makes you wonder what happened. Less forgiving.
- 2:38It was a complex coding assignment was evaluated
- 2:40using codecs. Using codecs kept the evaluation
- 2:43entirely neutral and fair. Exactly. And GLM 5
- 2:47.2 actually did very well here. It handled the
- 2:50bulk of the task perfectly. But. Opus caught
- 2:52one very subtle edge case that GLM missed entirely.
- 2:57tricky database values, it caught the subtle
- 2:59difference between true and one, or one and one
- 3:03point now. Oh, wow. If you have ever stared at
- 3:04a broken database at two in the morning, you
- 3:06know exactly why that matters. Yeah, bugs always
- 3:09live in those tiny details. It feels like a very
- 3:11natural division of labor. GLM is kind of like
- 3:14the fast junior developer. It writes the bulk
- 3:16of the code very quickly. And Opus is the senior
- 3:20architect catching the tricky database bugs before
- 3:22they launch. That is a perfect analogy for how
- 3:25this works. GLN is fantastic for fast implementation
- 3:29and initial scaffolding. Right. And OPUS is necessary
- 3:32for careful reasoning and tricky edge case handling.
- 3:36Exactly. I have to be honest here. I still wrestle
- 3:39with prompt drift myself. beat. You know, you
- 3:42give an AI a long list of instructions. And by
- 3:45step three, it completely forgets step one. So
- 3:48watching an open source model hold its focus
- 3:50over complex instructions is deeply impressive
- 3:52to me. It really is an incredible leap forward.
- 3:54Right. But there's a specific speed quirk we
- 3:56definitely should mention. Right. GLM is not
- 3:58always the faster model. Yeah. On a highly creative
- 4:01HTML task, GLM took 35 minutes. Opus finished
- 4:05that exact same task in only 11 minutes. The
- 4:08rule here seems to revolve entirely around reasoning.
- 4:11Exactly. The more reasoning required by the prompt,
- 4:13the slower GLM feels in practice. Execution -heavy
- 4:16tasks are incredibly fast. Planning and creative
- 4:19taste just take much longer. They both succeeded
- 4:22on the creative task, though. GLM built an interactive
- 4:26page called the Anatomy of Attention. Right.
- 4:28It featured moving background elements and token
- 4:30visuals. and Opus built the life of a Death Star.
- 4:35It was a beautifully structured timeline -style
- 4:37page. Both were excellent one -shot results.
- 4:40The final test here was deep research. This utilized
- 4:43a storm -style workflow. For anyone unfamiliar,
- 4:47that just means multiple sub -agents work together
- 4:49using different personas. Right. They compiled
- 4:51a very rich, highly detailed HTML report. It
- 4:55combined different expert lenses and actively
- 4:58challenged its own assumptions. And GLM managed
- 5:01this complex agent workflow beautifully. The
- 5:04significantly lower cost changes the math entirely
- 5:06for us here. Running multiple autonomous agents
- 5:09suddenly becomes justifiable for daily work.
- 5:11Yeah, you can test 10 different angles without
- 5:13burning through your cash. But wait, if GLM struggles
- 5:16with heavy reasoning, doesn't that make it too
- 5:17risky for professional work? Well, not necessarily.
- 5:21It is really about matching the model to the
- 5:23risk profile. You have to match the specific
- 5:25tool to the specific task. Right. Scaffolding
- 5:28a basic website layout carries very low risk.
- 5:31Exactly. But migrating a production database
- 5:33carries extremely high risk. You always use the
- 5:37heavy reasoning model for the high risk work.
- 5:38So you don't compare final answers, you compare
- 5:41your acceptable risk for the task. Precisely.
- 5:43Which leads us to a much broader mindset shift
- 5:46we need to discuss. Yeah. The real skill today
- 5:49is knowing exactly when to use GLM 5 .2. AI work
- 5:53is not a simple contest anymore. It is not about
- 5:55the best model wins mentality. No, it is a multi
- 5:59-step, highly iterative process. Most work involves
- 6:03researching, drafting, and continuous testing.
- 6:05Then you edit, make decisions, and finally ship
- 6:07the product. Each of those steps demands a completely
- 6:10different level of intelligence. Exactly. Let's
- 6:12break down the 80 % work first. This is the natural
- 6:15domain of GLM 5 .2. It easily handles first drafts
- 6:19and gathering initial research. It does basic
- 6:21web design and cleans up your messy notes. Right.
- 6:24It generates your initial options for a project.
- 6:27Cost matters immensely here because you are doing
- 6:29constant iterative testing. You are exploring
- 6:32different ideas. You don't want maximum pricing
- 6:34when you are just sketching a rough draft. Exactly.
- 6:37Then we have the remaining 10 to 20 percent work.
- 6:39This is the domain of OPUS 4 .8. This is the
- 6:42heavy thinking. Final reasoning, edge case review,
- 6:45and high -risk coding tasks. The system context
- 6:48is crucial here, too. The harness really matters.
- 6:51Right. For clarity, a harness is just a digital
- 6:53workspace where the AI can use tools. Cloud Code
- 6:56provides that powerful harness for us. It lets
- 6:59the model read local files and actively run terminal
- 7:01commands. So GLM is not just a cheap OPUS replacement.
- 7:05It is a cheaper worker operating inside the exact
- 7:08same system. Mm -hmm. Beat. But aren't we just
- 7:11complicating things by juggling multiple workers
- 7:13for a single project? It might seem that way
- 7:16at first glance. But using one expensive model
- 7:18for basic sorting tasks burns cash needlessly.
- 7:22Right. Splitting the work optimizes both your
- 7:24budget and the applied brain power. You get significantly
- 7:26better efficiency across the board. Let the cheap
- 7:29model gather the lumber. Let the expensive model
- 7:31build the house. Exactly. It is about working
- 7:34much smarter within the environment you already
- 7:36use every day. Placeholder for sponsor reads
- 7:39skip promotional text from newsletter, use provided
- 7:42sponsor copy. Hearing about those massive price
- 7:45differences makes me want to try this immediately.
- 7:48Yeah. But whenever we talk about open source
- 7:50routing, it usually involves spinning up servers
- 7:52or Docker containers. Right. Is this actually
- 7:56feasible for a normal user to set up? It absolutely
- 8:00is. You are not learning a brand new tool here.
- 8:03You are just routing the model call to a different
- 8:05location. You simply route it to z .ai instead
- 8:08of Anthropix server. Exactly. Step one is going
- 8:10to the z .ai API console online. You can actually
- 8:14test the model out right there first. They have
- 8:163D generation tools and small mini games available.
- 8:19It gives you a great feel for how the model responds.
- 8:22Yeah. Then you choose your preferred billing
- 8:24method. You can pay per token or choose a set
- 8:26monthly plan. Those monthly plans run roughly
- 8:29$16, $64, or $144. There's some really great
- 8:34practical advice on this point. Keep your Claude
- 8:37plan active and just add z .ai on the side. Yeah,
- 8:40you don't have to choose just one platform. You
- 8:42use both of them together. Next, you generate
- 8:44a secure API key inside the console. This brings
- 8:47us to editing a specific local file. It is called
- 8:50settings .local .json. You find this file sitting
- 8:55inside the .clod folder on your machine. What
- 8:58is brilliant here is that you aren't installing
- 9:00heavy new software. Clod code is already looking
- 9:03for a brain at a specific web address. Right.
- 9:06All you were doing in that settings file is changing
- 9:08the address book. You pointed away from Anthropic
- 9:11servers. You route the base URL directly to z
- 9:14.ai instead. You leave the Anthropic API key
- 9:18completely blank, and you put your new z .ai
- 9:21key in as the auth token. Finally, you set the
- 9:24specific model name you want to use. It is surprisingly
- 9:26simple. But there is a genius trick we need to
- 9:29highlight here. You create two entirely separate
- 9:31folders on your local machine. Yeah, this is
- 9:33incredibly clever. One folder is named slash
- 9:36gm. The other folder is named slash opus. You
- 9:39put the custom routing configuration file only
- 9:41inside the slash glm folder. Wait, so just by
- 9:45changing directories in your terminal, you instantly
- 9:48switch the brain powering your workspace, zero
- 9:51friction. That is brilliant. You open clod code
- 9:54in the glm folder for your rough drafts. You
- 9:56open it in the opus folder for your final code
- 9:58reviews. Exactly. But I have to ask a security
- 10:01question here. If I put my API key in a local
- 10:04JSON file, isn't there a risk I accidentally
- 10:07share it? Oh, there is absolutely a major risk
- 10:09there. That is why you must be extremely careful
- 10:12with this file. Right. Treat that API key exactly
- 10:15like a banking password. Keep it completely out
- 10:17of public repos or shared team screenshots. Treat
- 10:21the key like a password. Never push your local
- 10:23settings to the public. Exactly. Keep it secure
- 10:25and the entire workflow remains brilliant and
- 10:28safe. Now, we really need to look at the bigger
- 10:30picture here. Why does GLM 5 .2 make open source
- 10:34AI models impossible to ignore right now? Well,
- 10:37open source is finally practical for the daily
- 10:40professional workday. It is no longer just a
- 10:42theoretical weekend project for developers. The
- 10:45massive scale of this model is truly staggering
- 10:47to think about. GLM 5 .2 operates with around
- 10:51753 to 756 billion parameters. Yeah. Let's pause
- 10:57on that number for a second. That is far too
- 10:59massive to run locally on a normal computer.
- 11:02You would need serious, incredibly expensive
- 11:04server hardware in your house. That is exactly
- 11:07why API renting through providers is necessary.
- 11:10It is the pragmatic middle ground for users right
- 11:12now. You get all the massive power without the
- 11:15massive infrastructure headache. This brings
- 11:17us to the underlying cost of tokens. Tokens are
- 11:20just tiny puzzle pieces of text. It is kind of
- 11:23like stacking logo blocks of data. Exactly. The
- 11:26AI processes these tiny pieces to understand
- 11:28and generate language. When you look closely
- 11:31at the token pricing, the math is undeniable.
- 11:33Right. Opus costs $5 in and $25 out per 1 million
- 11:37tokens. GLM sits at roughly $1 .40 in and $4
- 11:42.40 out. Two secs silence. Whoa. I mean, imagine
- 11:46scaling to a billion queries without bankrupt.
- 11:48yourself. That five times price difference changes
- 11:52everything about how we build software. It is
- 11:55especially critical for complex agent workflows.
- 11:58These autonomous models constantly read folders,
- 12:01revise their own work, and call subagents. They
- 12:04burn tokens invisibly in the background while
- 12:06you drink your coffee. Exactly. Significantly
- 12:09lower prices encourage you to experiment freely
- 12:12without watching the meter. There are also performance
- 12:14benchmarks to consider. GLM actively beats GPT
- 12:185 .5 and Opus 4 .8 on some specific software
- 12:22benchmarks. Right, but there is a very crucial
- 12:25point about those numbers. Benchmarks are just
- 12:27a signal. They simply tell you a model is worth
- 12:29testing yourself. Daily usefulness on your own
- 12:32specific files is the only real test. Yeah, how
- 12:35does it handle your unique code and your messy
- 12:37personal notes? There is also a major strategic
- 12:40advantage we should point out here. It acts as
- 12:42a necessary hedge. Closed platforms change their
- 12:45rules and their pricing structures all the time.
- 12:47A vital feature can suddenly move behind an expensive
- 12:51paywall tomorrow. Learning how to route open
- 12:53source models protects your daily workflow from
- 12:55those sudden shifts. But think about this dependency
- 12:58for a moment. If open source models are this
- 13:01huge, won't we always be dependent on cloud providers
- 13:04to run them anyway? Yes. For now, that is the
- 13:07reality. Yes. API rental is the necessary middle
- 13:10ground today. Right. But it still fundamentally
- 13:13breaks the monopoly. You aren't relying entirely
- 13:16on one single closed ecosystems rules anymore.
- 13:20It's about having a backup plan when the closed
- 13:22platforms change their rules. Exactly. You maintain
- 13:25your professional options and your creative freedom.
- 13:28So let's bring all of these different pieces
- 13:30together for you. We are looking at a fundamental
- 13:32shift in how we approach knowledge work. The
- 13:34era of treating open source models as clunky
- 13:37budget alternatives is officially over. The true
- 13:40modern skill isn't finding the one perfect AI
- 13:43to do everything. You are essentially becoming
- 13:46an orchestrator of multiple minds. Right. You
- 13:48have to know exactly when to deploy a fast, cheap
- 13:50worker. You use GLM 5 .2 for the heavy, repetitive
- 13:54lifting and the rough drafting. And you must
- 13:56know exactly when to bring in the expensive protectionist.
- 13:59You call on Opus 4 .8 to close the deal and ensure
- 14:01total precision. It is a beautiful synergy when
- 14:04you set it up correctly. We want you to look
- 14:06at your own daily tasks today. Identify the 80
- 14:10% work you are currently doing manually. Where
- 14:12are you overpaying for AI compute right now?
- 14:15Where could a fast, highly capable open source
- 14:18model handle the drafting and sorting for you?
- 14:22Finding that exact balance will completely change
- 14:24your daily productivity. Thank you for joining
- 14:27us on this deep dive today. We always appreciate
- 14:30you spending your valuable time with us. It has
- 14:32been a truly fantastic exploration today. We
- 14:35want to leave you with one final provocative
- 14:37thought to mull over. If an open source model
- 14:40is already handling 80 % of the workflow today
- 14:43at a fraction of the cost, what happens to the
- 14:45value of closed models when open source naturally
- 14:48creeps up to cover 90 or 95 %? Does intelligence
- 14:51become essentially free?