Better than Claude and 10x cheaper?? GLM-4.6 is so good 🔗 **Source:** [Better than Claude and 10x cheaper?? GLM-4.6 is so good](https://www.youtube.com/watch?v=LJJnu0zfI6w) > [!note]+ Document Notes > tags: [AI-Models, Coding-Agents, Cost-Optimization] > author: [MISSING] > source: [YouTube] > reference: [https://www.youtube.com/watch?v=LJJnu0zfI6w] > date created: > > # 🏷️ Title: Better than Claude and 10x cheaper?? GLM-4.6 is so good > <!-- deep-read:sections start --> ## 🧰 Action Inbox - [H1] Draft a cost/quality brief for multi-model stacks — GLM-4.6 vs Sonnet 4 (video) / https://www.youtube.com/watch?v=LJJnu0zfI6w #articleseed - [H2] Pilot “planner → doer” model routing in dev tooling — React repo demo (video) / https://www.youtube.com/watch?v=LJJnu0zfI6w #xref - [H3] Create OSPO guidance on benchmark vs win-rate claims — use [H3][H5] as examples / https://www.youtube.com/watch?v=LJJnu0zfI6w #addendum - [H4] Build a WP Playground task list that cheaper models execute — reference planning pattern / https://www.youtube.com/watch?v=LJJnu0zfI6w #articleseed - [H5] Validate TPS claims with a reproducible harness — note 85 TPS remark / https://www.youtube.com/watch?v=LJJnu0zfI6w #articleseed - [H6] Share a “$3/mo coder” explainer with The WPOCC fellows — cost math and caveats / https://www.youtube.com/watch?v=LJJnu0zfI6w #share - [H7] Add a FAIR.pm note on transparency norms — publish eval sets like ZI example / https://www.youtube.com/watch?v=LJJnu0zfI6w #articleseed - [H8] Prepare DX checklists for multi-file coherence tests — compare agents on the same repo / https://www.youtube.com/watch?v=LJJnu0zfI6w #xref - [H9] Draft governance language on tool selection bias — avoid “dead model/company” claims without data / https://www.youtube.com/watch?v=LJJnu0zfI6w #aphorisms - [H10] Record a short demo: planner=Claude, doers=GLM-4.6 — measure cost and latency / https://www.youtube.com/watch?v=LJJnu0zfI6w #articleseed - [H11] Notify WordPress training team about cheaper agent options — propose sandbox labs / https://www.youtube.com/watch?v=LJJnu0zfI6w #share - [H12] Cross-reference community write-ups on GLM-4.6 reliability — add external sources later / https://www.youtube.com/watch?v=LJJnu0zfI6w #xref ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - Include links cited in or relevant to the article. ## 📝 Summary - The video reviews GLM-4.6’s coding and reasoning performance. - The host compares price and quality to Claude Sonnet 4. - Tooling and agent orchestration strongly affect results. - Reported win-rates and TPS look promising, but require replication. - Bottom line: GLM-4.6 seems cost-effective for routine coding tasks. ## 📖 Main Content - GLM-4.6 claims higher win-rates versus Sonnet 4 in pairwise judgments. See [H3]. - The model emphasizes token efficiency and longer context. See [H7]. - The host advocates a planner/doer setup: expensive model plans; cheaper model codes. See [H4][H10]. - Tool maturity limits outcomes; multi-file coherence matters. See [H8]. - Transparency improves trust when labs share eval data and trajectories. See [H7]. - Uncertainties remain: host-reported TPS, dataset quality, and generalization. See [H5]. ### 🗒️ Notes #### ⏱️ [00:03](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=3) Context and positioning ### [H1] GLM-4.6 positions as cheap alternative > [!quote] [Video](https://www.youtube.com/watch?v=LJJnu0zfI6w) by an unnamed author via [[YouTube]]. Published [MISSING]. Accessed . > “we might finally have a cheap, reasonable alternative to using cloud models.” Short rationale: Sets thesis and value frame. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=3) #articleseed ID:rw_[MISSING] #### ⏱️ [06:39](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=399) Head-to-head result ### [H2] Reported win-rate approaches parity with Sonnet 4 > [!quote] [Video](https://www.youtube.com/watch?v=LJJnu0zfI6w) > “4.6 is winning 48.6% of the time against Sonnet 4.” Rationale: Highlights near-parity; invites replication. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=399) #xref ID:rw_[MISSING] #### ⏱️ [04:36](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=276) Agentic features ### [H3] GLM-4.6 improves tool-use and reasoning > [!quote] [Video](https://www.youtube.com/watch?v=LJJnu0zfI6w) > “supports tool use during reasoning and inference… a more capable agent.” Rationale: Positions model for agent workflows. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=276) #xref ID:rw_[MISSING] #### ⏱️ [13:32](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=812) Orchestration idea ### [H4] Planner/doer routing reduces cost > [!quote] [Video](https://www.youtube.com/watch?v=LJJnu0zfI6w) > “top level planning first… sub agents using a different cheaper model.” Rationale: Suggests architecture for WordPress dev workflows. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=812) #articleseed ID:rw_[MISSING] #### ⏱️ [08:38](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=518) Throughput claim ### [H5] ZI API reports high TPS > [!quote] [Video](https://www.youtube.com/watch?v=LJJnu0zfI6w) > “The official ZI API is doing it at 85 TPS.” Rationale: Throughput affects developer ergonomics and CI times. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=518) #articleseed ID:rw_[MISSING] #### ⏱️ [09:39](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=579) Pricing ### [H6] Coding plan advertises $3 per month > [!quote] [Video](https://www.youtube.com/watch?v=LJJnu0zfI6w) > “At $3 a month for their GLM coding plan…” Rationale: Enables broader contributor access; watch usage caps. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=579) #share ID:rw_[MISSING] #### ⏱️ [04:03](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=243) Context length and tokens ### [H7] Longer context and fewer tokens per task > [!quote] [Video](https://www.youtube.com/watch?v=LJJnu0zfI6w) > “Longer context windows up to 200k tokens.” Rationale: Large diffs and multi-file edits become practical. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=243) #xref ID:rw_[MISSING] #### ⏱️ [09:26](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=566) Multi-file coherence ### [H8] Model maintains multi-file coherence better > [!quote] [Video](https://www.youtube.com/watch?v=LJJnu0zfI6w) > “maintains coherence across multi-file operations…” Rationale: Key for plugin and theme repos. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=566) #xref ID:rw_[MISSING] #### ⏱️ [15:20](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=920) Tooling gap ### [H9] Current tools lack flexible model routing > [!quote] [Video](https://www.youtube.com/watch?v=LJJnu0zfI6w) > “none of these tools have this functionality.” Rationale: Opportunity for IDE integrations and CI bots. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=920) #articleseed ID:rw_[MISSING] #### ⏱️ [10:42](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=642) Transparency practice ### [H10] Lab shares test questions and trajectories > [!quote] [Video](https://www.youtube.com/watch?v=LJJnu0zfI6w) > “They published all their test questions and agent trajectories…” Rationale: Supports FAIR.pm style reproducibility norms. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=642) #articleseed ID:rw_[MISSING] #### ⏱️ [18:49](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=1129) Practical verdict ### [H11] Not daily driver yet for the host > [!quote] [Video](https://www.youtube.com/watch?v=LJJnu0zfI6w) > “Will I be using this model every day? Not yet.” Rationale: Temper expectations; prioritize fit-for-purpose tasks. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=1129) #addendum ID:rw_[MISSING] #### ⏱️ [19:13](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=1153) Future direction ### [H12] Cheaper sub-agents under premium planners > [!quote] [Video](https://www.youtube.com/watch?v=LJJnu0zfI6w) > “I want to see GPT5 spinning up sub agents using these cheaper models.” Rationale: Guides future WordPress tooling experiments. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=1153) #articleseed ID:rw_[MISSING] ## 📚 References - Video: GLM-4.6 review and demos — YouTube: https://www.youtube.com/watch?v=LJJnu0zfI6w - Note: Official GLM-4.6 release notes link not provided in transcript. Add when available. ## 🗣️ Paraphrase - The presenter reviews GLM-4.6, emphasizing cost, speed, and coding skill. - They propose using expensive models for planning and GLM-4.6 for implementation. ## 📰 Newsletter - GLM-4.6 looks like a thrifty coding partner this week. - Reported win-rates approach Claude Sonnet 4 at much lower cost. - We will test planner/doer routing in WordPress plugin repos. - Want to help? Join a Playground lab and measure coherence on multi-file edits. - Watch: GLM-4.6 review → https://www.youtube.com/watch?v=LJJnu0zfI6w ## ✅ Action Items #### DevRel - Publish a planner/doer tutorial using a WordPress starter. - Host office hours on cost-aware agent stacks. - Ship minimal examples for multi-file refactors. #### OSPO - Add guidance on model selection, metrics, and replication. - Define acceptance criteria for eval transparency. - Track cost per merged PR as a KPI. #### Community - Run contributor sprints with cheaper doer models. - Offer pairing on agent setups for new maintainers. - Share outcomes and configs in public docs. #### WordPress - Core: prototype coherent refactors with agent checks. - Playground: add labs for planner/doer orchestration. - Documentation: record “model routing” patterns and pitfalls. - Training: create lessons on token budgets and context. - Meta: instrument CI to log model cost and TPS. #### The WPOCC - Pilot microgrants for transparent, replicable evals. - Fund multi-model routing experiments in community repos. #### FAIR - Propose a standardized eval bundle and release checklist. - Draft covenant for publishing prompts, questions, and traces. ## 🧪 Hypotheses - GLM-4.6 reduces cost per accepted patch by ≥50% versus premium models. - Planner/doer routing preserves code quality versus single-model runs. - Multi-file coherence improves contributor throughput in plugin repos. - Transparent eval bundles raise adoption among maintainers. ### 📊 Methodology - A/B tasks across branches with identical repos and tickets. - Measure latency, TPS, tokens, and reviewer change requests. - Blind review merged diffs for regressions and style mismatches. - Publish prompts, runs, and traces for replication. ### ❓ Questions - Which WordPress tasks benefit most from cheaper doers? - How stable are win-rates across domains and languages? - What context size actually shifts coherence in practice? ### 🔮 Future Research - Agent routing policies tuned by repo metrics. - Self-healing runs that retry with stronger models. - Cost-aware planning that budgets tokens before runs. ### 💡 Implications - Teams can reserve premium models for hard planning. - Community projects can stretch funds with cheaper execution. - Transparent evals become a norm for vendor claims. ### 🧠 Learning Styles Assessment - Reading level estimate: Flesch-Kincaid around grade 10–11 (approximate). #### 👀 Visual learners - Present: UI demos and repo steps. - Improve: flowcharts for planner/doer routing. #### 🔊 Aural learners - Present: narrated walkthrough. - Improve: 3-minute audio recap and key takeaways. #### 🗣️ Verbal learners - Present: definitions for TPS, win-rate, context. - Improve: compare orchestration metaphors to teams. #### 🧩 Kinesthetic learners - Present: runnable repo demos. - Improve: Sandbox tasks with checklists. #### 🧠 Logical learners - Present: metrics and hypotheses. - Improve: decision trees for model routing. #### 🧑‍🤝‍🧑 Social learners - Present: collaboration prompts. - Improve: cohort reviews of diffs. #### 🧘‍♀️ Solitary learners - Present: self-paced labs. - Improve: reflection forms after runs. #### 🌿 Naturalistic learners - Present: none explicit. - Improve: energy and cost footprint notes. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: clear headings in this note. - Improve: add alt text to demo images. #### Keyboard navigation - Present: predictable steps. - Improve: ensure focus states in demo UIs. #### Contrast and readability - Present: high-contrast code blocks. - Improve: verify 4.5:1 in demo themes. #### Text resizing and reflow - Present: responsive docs. - Improve: honor reduced-motion preferences. #### Media and captions - Present: YouTube captions exist. - Improve: downloadable transcript and chapter markers. #### Forms and interactions - Present: N/A for video. - Improve: feedback form with clear labels. #### Standards reference - Map gaps to WCAG 2.2 AA and document fixes. ### 🌍 DEIB Assessment #### Cultural breadth - Present: global developer framing. - Improve: multi-language captions and subtitles. #### Inclusive language - Present: mostly neutral phrasing. - Improve: avoid disparaging remarks about vendors. #### Neurodiversity support - Present: chunked sections and lists. - Improve: consistent icon+text labels. #### Representation - Present: single narrator format. - Improve: feature diverse guest voices. #### Psychological safety - Present: constructive critiques. - Improve: include moderation guidelines in comments. ### ✅ Logical Validity #### Strong reasoning - Inductive comparison using win-rates ([H2]) supports parity claims. - Causal claim that orchestration reduces cost aligns with architecture logic ([H4]). - Transparency improving trust rests on evidence sharing ([H10]). #### Fallacies - Risk of hasty generalization from limited benchmarks ([H2]). - Fix: add multi-domain replications and preregistered tests. - Possible appeal to novelty regarding tool-use features ([H3]). - Fix: compare against baselines with ablations. #### Tone of voice - Energetic, pragmatic, occasionally blunt. - Encourages experimentation while noting limits ([H11]). ### 🛡️ DARVO Analysis - Deny: no denial patterns observed. - Attack: mild vendor jabs appear; keep focus on data. - Reverse victim/offender: not present. ## 🎯 Conclusion - GLM-4.6 appears strong for cost-sensitive coding tasks. - Use premium models for planning; route execution to cheaper agents. - Replicate win-rates and TPS before policy changes. - Share traces and results to build community trust. ## 🏷️ Tag Coverage #articleseed: 5 • #xref: 3 • #share: 2 • #bmf: 0 • #addendum: 1 • #aphorisms: 1 ## 🧱 Refactor Appendix (append after 🎯 Conclusion) ### 📚 Literature Split Pad - Purpose: Stage Konik CEE units that become per-source Literature notes. - Split level: **H3**. - Include verbatim quotes under 25 words and a [view highlight] link. ### [H1] GLM-4.6 positions as cheap alternative > [!quote] [Better than Claude and 10x cheaper?? GLM-4.6 is so good](https://www.youtube.com/watch?v=LJJnu0zfI6w) by an unnamed author via [[YouTube]]. Published [MISSING]. Accessed . > “we might finally have a cheap, reasonable alternative to using cloud models.” Two lines: context and use. Cite replication steps and costs. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=3) #xref #articleseed ID:rw_[MISSING] ### [H2] Reported win-rate approaches parity with Sonnet 4 > [!quote] [Better than Claude and 10x cheaper?? GLM-4.6 is so good](https://www.youtube.com/watch?v=LJJnu0zfI6w) > “4.6 is winning 48.6% of the time against Sonnet 4.” Two lines: plan replication with WordPress tasks. [view highlight](https://www.youtube.com/watch?v=LJJnu0zfI6w&t=399) #xref ID:rw_[MISSING] --- ### 🔖 Atomic Split Pad (Zettel Candidates) - Purpose: Promote reusable claims into Atomic notes. - Split level: **H4**. - Each item: one idea, present tense, specific, testable. #### [Z1] Planner/doer routing reduces model spend - Use premium planner, cheap doers; track quality. - Links: [[Model-Orchestration]] [[Agent-Routing]] #### [Z2] Transparent eval bundles increase trust - Publish prompts, traces, and outcomes for reuse. - Links: [[FAIR.pm-Principles]] [[Evaluation-Playbooks]] --- ### ✏️ Fleeting Queue (Extract by Selection) - Purpose: Capture quick actions or sparks for Fleeting notes. - Extraction: **Select one line** → “Refactor: Extract selection to new note (content only).” - Keep each line short and self-contained. - [F1] Build a small plugin repo to A/B planner/doer routing — context/https://www.youtube.com/watch?v=LJJnu0zfI6w #fleeting #inbox - [F2] Draft a template for publishing eval traces — context/https://www.youtube.com/watch?v=LJJnu0zfI6w #fleeting #inbox - [F3] Prepare a WP Playground lab on multi-file coherence — context/https://www.youtube.com/watch?v=LJJnu0zfI6w #fleeting #inbox --- <!-- deep-read:sections end --> ## 📺 Video  --- # 📄 Full Transcript *Auto-generated from YouTube captions · timestamps roughly every 30 seconds · [watch on YouTube](https://www.youtube.com/watch?v=LJJnu0zfI6w)* **[00:00]** It's been a while since we had a new openweight model that was really worth diving into. It just kind of feels like the Frontier Labs are the ones pushing things right now. And historically, they have been. They have a ton of money and resources and all the other things they need, but they also charge accordingly. I miss the days where Deep Seek came in and shook up the whole industry. And while we're not in one of those moments right now, it does feel like we're really close because the ZI have been cooking. I am genuinely really impressed with what they've been doing on many different levels. On one hand, they're an incredible lab making models that are scoring super well on various **[00:31]** benchmarks. But on the other hand, they seem to understand and care about product a lot more than the other Chinese labs. While they're putting this out, Deepseek is putting out yet another weird version of Deepseek V3 focused on their sparse attention methods. Deepseek's a research lab through and through. They're so bad at software development that they inspired me to make T3 chat because I hated using their chat interface so much. Whereas with ZI, the first link's the API, then the next link is their subscriptions, which you might remember previously when I did the video on how to make cloud code way cheaper. With ZAI, the model just got **[01:03]** way better than it was then, and now it's beating out legit things. Like, it has a good win rate against Cloud Sonnet 4, which is kind of crazy. There is so much to dive into here. I am genuinely hyped. Zi has been cooking and we might finally have a cheap, reasonable alternative to using cloud models. According to kilo code, 48.6% win rate against claude sonnet 4. So it's like neck andneck with sonnet 4 for oneif or less the price. For output tokens, it's like one. It's insane. This is a really **[01:36]** cool model with a lot to dig into. I can't wait to try it myself and show you guys more. Normally I complain about token costs, but it's cheap enough. I don't have much to worry about here. That said, we do have bills to pay. Quick word from today's sponsor and then we'll dive right in. Adding AI to your app has never been easier, but making it useful is way harder than I would have ever imagined. If you're trying to make it so users can access things from other services like, you know, Google Drive, Notion, Slack, or even Salesforce, getting the integration setup right is nearly impossible. And then actually making it so the AI can find what it's **[02:07]** looking for from those sources without just bloating its context and getting super dumb and expensive. You're not going to pull it off. Believe me, I've tried. And that's how I stumbled on Raggy. These guys really get it. They are rag as a service, but it goes so much deeper than that. You know that service plaid that lets you sign in with your bank information to any other place when you want to link your bank account to something. You can almost think of raggi like that, but for your data. So, if you're building a service and you want your users to be able to index and bring in their Google Drive information, their Slack messages, or basically any **[02:38]** source that you could ever imagine, Raggy will make it easy for them to sign in, index the data for you, and now your AI has the tools it needs and the data it needs to answer questions about all of that. And the level of depth they went to with this is genuinely crazy. Things get really crazy when you start adding audio and video, which they will not only index for you, but they will find spots like actual timestamps and stream that back to you. Here's an example where Raggy was used to find scenes with hard hats from a random set of videos, and it was able to stream in the exact timestamp and location with metadata and information where this **[03:09]** moment happened in the video. Do you know how much work it is to build this type of stuff yourself? I have. It's not fun. And the fact that they are so cheap makes it even crazier. All of their tiers offer unlimited retrievalss, which is crazy. The actual ingest processing is pretty good, too. Only 2 cents per page of data and 5 cents if you want to do the high-res version of that. Two cents per minute of video is insane for processing, too. Like, there is so much stuff you can build with this and it makes me want to add integrations on T3 Chat like right now. Seriously, this platform is so cool. The team's been awesome to work with. They are huge fans **[03:41]** of us and everything we do here. If you're looking for a product that solves these problems and deeply understands our expectations as fullstack TypeScript nerds, there's no better place to look than Raggy. Check them out now at soyv.link/raggy. GLM 4.6 advanced agentic reasoning and coding capabilities. Today we're releasing the latest version of our flagship model. Compared with 4.5, this generation brings several key improvements. Longer context windows up to 200k tokens. Superior coding performance. The model achieves higher scores on code benchmarks and demonstrates better real world **[04:11]** performance in apps like Claude Code, client, R code, and kilo code. You might be confused about why they put cloud code here. It's because they published a pretty solid guide instead of scripts so that you can use cloud code with their models. Since you can set cloud code to hit a different endpoint, you can use it to hit their endpoint and then they crafted their endpoint to be as close as possible to the official cloud code one. So you could use it with cloud code. It's also much better at reasoning and it supports tool use during reasoning and inference which leads to stronger overall capabilities. Also it's a more **[04:41]** capable agent. GLM 4.6 exhibits stronger performance in tool use and searchbased agents and integrates more effectively within agent frameworks and is better at writing too. Cool. Let's look at these benchmarks in AM. It's beating out Sonnet 4.5 even which is crazy. GPQA it's losing to Sonnet 5 or 4.5 but beating out Sonnet 4. Live Codebench it's crushing. HLE it's crushing. T2 Bench it's doing pretty well. Better than four worse than 4.5. You get the idea. The only thing they're really losing against anthropic models in is S **[05:12]** swbench verified but it's still [ __ ] neck and neck. They're doing very good here. Sadly we don't have my favorite set of benchmarks. the artificial analysis charts for this just yet. But we can see that 4.5 did perform pretty well before. There's so much junk in here that I don't give a [ __ ] about. Yeah, it's been so long since a law model mattered at all. I'm getting rid of Gemini. They haven't been good for a bit. Excited for Gemini to catch up, but right now is not the time. Opus is effectively a dead model series now. Mestral is effectively **[05:44]** a dead company now. R10528. Yeah, not really useful at this point. It is interesting to see that with Deepseek V 3.1 to 3.2 that 3.2 performs slightly worse, but it's also way cheaper and easier to run due to their new weird techniques. Uh, now I have a more reasonable set of things. And you'll see GLM 4.5 doesn't necessarily perform great when you compare it to like GPT5 on minimal or Kimmy K2, but it's performing decently. And if we look at code benchmarks like we have here, 4.5 is up there with 3.1 beats out quenb **[06:16]** 3 max beats out gro 4 fast even but it's still not close to the best like GPT5 codeex and whatnot. 4.6 appears to be meaningful improvements compared to 4.5 according to their own numbers. Significant depending on the bench like live codebench it's crushing their previous implementation. But what really matters is win rates against other models. So here we have 4.6 versus Cloud Sonnet 4, which is asking people or an AI which of these solutions is better. 4.6 is winning 48.6% of the time against **[06:50]** Sonnet 4. Sonnet beats GLM 4.6 42% and then there's a 9.5% draw rate. So GLM 4.6 is actually performing better than Cloud Sonnet 4 here. It's also performing better than 4.5 for GLM models. These numbers are always hard to read into because it's kind of crazy when you think about it that 4.5 of a model is beating 4.6 sometimes. That is the nature of things. It's [ __ ] crushing Kimmy though. The one of the coolest parts though, I'm talking more about this token usage. It is using way fewer tokens per request than Deepseek **[07:20]** is or Kimmy is. And it's gone down from their previous model. It's nice to see Labs focusing on their token overusage and trying to trim it. That's all cool, but I want to actually try this out. I checked various different code agents. Uh, Open Code doesn't have it yet. I don't feel like breaking and resetting up my Cloud Code again, as fun as it is. What I really want to test this out with is Kilo Code. It's already built in. Kilo Code is a sponsor. They're not sponsoring this video. They're just general channel sponsors. I'm excited to see how it behaves with the latest model **[07:51]** here. gave it the same prompt I was using before, which is demonstrating the new activity API, which is a fun challenge because this isn't information they would have trained on because it's a new API just added to React. So, it has to be able to search the web to find more information about this. It's cool to see things update like this. I was just using an ancient version of Kilo and it half failed. Now that I'm on the latest version, it looks way better. It's calling tools way better. It's using a to-do list. Oh, this is actually really good. Kilo is not sponsoring this video. They've just sponsored them in the past. I don't even think we have any **[08:21]** pending sponsorship stuff right now, but it's good. Let it run the command. This kind of just feels like faster cloud so far in a lot of ways. A new directory. It's writing a bunch of code. It's [ __ ] fast. What are the TPS numbers they're seeing for this on open router. The official ZI API is doing it at 85 TPS, which makes it the fastest of this tier of model by quite a bit. Deep infra is only hitting 30 TPS and their uptime is garbage. Parasale is way slower at 20-ish TPS and their uptime is fine. So, **[08:53]** it seems like ZI is the best host by far right now. I'm excited to see others hosting it because it is an openweight model. And I honestly I can already say just seeing this, this model is definitely underrated right now. I had a couple people I trust DMing me saying that this model's cracked and I really need to check it out. I don't just do these one-off openweight model videos otherwise because there's so many of them, but this one does feel different. This is hitting harder than I expected. Bad tool call. What happened here? It's running things in parallel right here by the looks of it. It's editing multiple files at once. While that's running, **[09:23]** there are other things I wanted to cover. Again, kilo code covered this well here. The real story here isn't benchmarks. It's practical performance. 4.6 maintains coherence across multifile operations where other models typically hallucinate functions or lose context. At $3 a month for their GLM coding plan, it's cheaper than your streaming service. Here's what smart orchestration looks like. You use cloud or GBD4 for architecture and complex planning. You wrote implementation work to GLM 4.6 and you save 50 to 100x on cost for routine coding tasks. Even if it handles just **[09:53]** 80% of your workload, the return on investment here is obvious. ZI also did something unusual here. They published all their test questions and agent trajectories on HuggingFace. You can verify their claims yourself. See the actual code that GLM 4.6 generated where it succeeded and where it struggled. Transparency is way better than marketing benchmarks. I totally agree. They have been really weirdly transparent and nice to work with. Like of the recent labs, ZI was one of the first to like DM me and make sure I had what I needed in order to do good **[10:24]** coverage. They went out of their way to like involve me early. They've been awesome to work with and chat with, and they've been super transparent about things, too. 6 months ago, Frontier Level Coding AI meant frontier level prices. Today, models like GLM 4.6 are closing the gap fast. From Opus losing relevance due to Sonnet 4.5 to GPT5 dropping at a relatively cheap price without some crazy expensive higher tier like we all expected to now GLM 4.6 heavily closing that gap and beating out quad 4 and GPT4 for way cheaper is **[10:56]** crazy. These are legitimate alternatives that handle most coding tasks at a fraction of the cost. Save the premium models for the truly complex things. Use efficient models for everything else. What is this video they're linking? I'm very curious. Oh, AI code king. Yeah, he actually was one of the ones who DM' me saying, "This model's different. You should really pay attention." I can't believe I'm not subscribed. I'm going to turn on notifications for him because his channel has been awesome. Also, congrats AI Code King on breaking 100K. Very welld deserved. Models that were unthinkable at this price point 6 months ago are available now for the cost of a **[11:26]** coffee. That changes the economics of AIS in development entirely. This won't be your favorite model overnight, but at $3 a month through their coding plan, it's pretty absurdly cheap. Anyway, let's see how it did. Here it is finished generating this. Let's run the code. And here we are. The demo it created. Seems like it's infinite looping a render. Kind of funny. I will tell it it's doing that and see how it fixes it. While that's going, I'm going to do a quick open code run using same model. **[11:59]** Back for my activity demo. I asked it to fix the constant rerendering bug and it did almost immediately. That's actually cool. It made a nice solid UI. Fell for the making a bunch of cards at the bottom. It has a very different style than I normally see from these models. I'm actually really excited now to see how this open code run goes using the model to do the the classic the fake image studio. Definitely not the fastest solution for this. I was hoping there would be better examples around of using **[12:30]** smarter models like Claude or GPT5 to solve problems at like a higher level and then have GLM 4.6 go in and do the majority of the coding tasks at a fraction of the cost. I'm not finding examples anywhere of people doing this. I know what it would look like, but I don't know any existing tools that have this implemented particularly well just yet, which is a shame because it'll be really cool once they do. What I'm imagining here is something along the lines of like the to-do list that we saw being created earlier when I ran this. Or even now, if we hop into my open code **[13:03]** runner, you can see if we go up far enough, we have the to-do list where it's going to initialize the project, set it up with the basic layout, create dark mode, and global theme design and implement. You can imagine a world where a lot of that work is done by a more expensive, fancier model like claude 4.5 sonnet or GPT5. It plans out the to-do list, maybe even describes a bit more in detail what each of those tasks are and what they look like. And then it spins up a coding agent to actually go in **[13:34]** under the hood and do those things. That's already kind of what happens. the to-do list is looked at and then a sub agent is spawned to go do each of those tasks and report back when it's done. Having those sub agents using a different model than the top level agent makes a ton of sense. Cloud code kind of does this the other way where they have a cheaper model that will go look for information and try to help the more expensive model know what to do, right? They'll use haiku for a bunch of analysis of data in the repo and then let sonnet do the work after. I could **[14:04]** see a future where we have another layer on top that is the expensive model figuring out what to do and the cheaper models go do the actual details. Super exciting stuff. It appears to be going much further into this demo app than others do. Like they're implementing zoom and things. Curious to see how far it goes and what the results end up looking like. We are like 5 plus minutes into this run now. So, it's not the fastest model. It puts out tokens pretty quick and it feels fast when you're watching it edit in your editor for that reason. But it also seems to go a lot further with the generation work it's **[14:34]** doing which can result in it not being the fastest solution but also going a bit further. Just found a fun bug in open code. This generating should be animating but apparently probably due to some layers between my ghosty implementation as well as with T-Mox. I'll never leave T-Mox. You'll pull it out of my cold dead hands. This animation doesn't happen unless I click. And every click, you'll see I get an updated frame. Kind of funny. What's going on, guys? I **[15:04]** want to see the output. It appears to be stuck. As I think about this, I'm imagining a UI that's a bit different from what we have today. Imagine you have your main agent model and something like cursor. Here we have 4.5 sonnet max mode. Maybe that's our default. It would be really cool if we could have another little drop down here or a setting somewhere that lets us pick a different model for the subtasks that are created once you do a run. When you tell it, I want this, go implement it, and it describes what it has in mind. Then it can have those sub aents run with a **[15:36]** different cheaper model to help reduce your costs a ton because running one of these big models over giant amounts of context over and over again can get really expensive. But if you have a good enough plan, the dumber models can go do it. You can almost think of this kind of like a traditional engineering team. You need somebody who knows things well enough to do the higher level planning to make sure every like task is cut properly and that the overall structure and direction of the project is in the right direction. And then you can have less aware engineers doing the implementation details of those given **[16:08]** tickets. It's always a good idea to have somebody who really gets [ __ ] do that top level planning first. And the idea of having sub selection of dumber models that are way cheaper to do these other tasks is really cool. And I wish more of these tools were built with that instead of the tool just using one model for everything for a given run. The more I think about it, the only tool I know of that is really good about using different models for different things, funny enough, is Claude Code. Because cloud models are so expensive, they had to take advantage of cheaper cla models like you know haik coup in order to make **[16:40]** the product viable at all. But as far as I know, none of these tools have this functionality. Even things like kilo code who are actively talking about this idea of like using different models for different parts of the work, they don't have this implemented yet. At least as far as I know. Let me check settings and see if they have anything for this. Be very surprised if they did. Yep, they don't. Apparently, it's spitting out invalid JSX. That's fun. Like, I genuinely feel as though the pain I'm experiencing right now is because the tools have not caught up for this. And **[17:11]** I'm seeing more why the major labs are putting so much effort into things like Cloud Code and Codeex because you get so much value out of owning that and making sure the experience with your model is just right. I've seen a lot of speculation around businesses building these things saying, "Well, you can just swap the model. It's not that big a deal. These tools aren't that complex." They shouldn't be that complex. But I'm seeing a lot of random [ __ ] like this happening that is causing problems for a lot of people. Cool. Supposedly, it's now done. And it just wants to test **[17:42]** the complete app. I'll go check out local host 3000. See what it did finally. Well, uh, this actually looks better than what I would normally get out of Claude. There are weird things like the header sucks, but this is like a starting point. It's actually pretty good. I like the gradient. I don't like the hover behavior. It's hilarious that it just it hit a client side exception. **[18:14]** Oh, this is just cuz it doesn't have in the next config for the mock image. Annoying, but happens. There's a ton of potential here. That much is clear. The benchmarks show it. The quality of that UI by default shows it. The hype that lots of other companies have around the model shows it especially well. Will I be using this model every day? Not yet. But would I use a tool that uses this model as one of the main solutions for these types of problems? Absolutely. I would love for the claude parody claim **[18:45]** here to not just be its quality of code in benchmarks can be similar to claude, but the quality of experience you get using it is too. There's absolutely the potential for it to get there, but I'm not seeing that potential just yet. Very excited to see where this all goes. There's more reason now than ever to have multiple models solving a problem in tandem together. The models don't know other models are doing different parts. They have no idea. All they have is text in, text out. I want to see more companies take advantage of this. I want to see GPT5 spinning up sub agents using **[19:17]** these cheaper models and see what the results look like. If somebody ends up building that tool and makes it easy for me to try, I would love to give it a shot and use it for future videos. But for now, I just have to stare at my open code terminal very slowly failing, running into tons of issues, and then 25 plus minutes later having an actual answer kind of gh. They're trying so hard to reduce costs while also increasing intelligence. And that's an important thing for us as an industry to strive for. I can't wait to see what people do with this model and I'm excited for the future of tooling people will build around models like it. But **[19:48]** for now, it's not quite what I would use. Let me know what you guys think.