Nate Herk | AI Automation 6/20/2026 16 min video 3 slides

Running GLM 5.2 Inside Claude Code: An Open Model at One-Fifth the Cost

#GLM 5.2 #Claude Code #open-source model #local model #Opus 4.8 #Anthropic

GLM 5.2 in Claude Code is Blowing My Mind

83,545 views 6 highlights

TL;DR · What you'll learn

1 After spending a full day running the open-source large model GLM 5.2 inside Claude Code's harness, the creator concludes that it is fast, far cheaper, and fits surprisingly well into the workflow. It even edited the intro of the video itself from raw footage.
2 Using a single prompt (`/goal`), it generated a 23-second video over 1 hour and 15 minutes while consuming roughly 357,000 tokens. Some tasks finished faster than with Opus, while others still favored Opus.
3 Unlike closed models such as ChatGPT and Claude, GLM is open source, though GLM 5.2 is still so large at 753 billion parameters that the creator cannot run it on personal hardware and ends up renting access online.
4 A full day of heavy coding costs about one-fifth as much as Opus 4.8. With Opus 4.8 priced at $5 for input and $25 for output, the gap in cost is a major reason people are excited.
5 The creator predicts a future where every company runs its own local models, and argues that Anthropic has already begun investing beyond the model itself, including services and forward-deployed engineers for enterprise integration.
6 The performance gap between closed and open models is still significant, but it is shrinking quickly, and the creator says watching that happen in real time is part of what makes the moment so interesting.

Read as slides

3 slides total

01 Slide 1 / 3

Watch at 00:00

Testing GLM 5.2 inside Claude Code's harness

After spending an entire day working with the open-source large model GLM 5.2 inside Claude Code, the creator's takeaway is that it is fast, dramatically cheaper, and a good fit for Claude Code's harness. As evidence, he says the intro viewers are watching was itself edited by GLM 5.2 from raw footage all the way to the final version.

In practice, a single prompt (`/goal`) generated a 23-second video, taking 1 hour and 15 minutes and consuming roughly 357,000 tokens. Some tasks finished faster with GLM, while others were clearly faster with Opus. His conclusion is that this is not an either-or choice, but a case for deciding which model fits each stage of the workflow.

02 Slide 2 / 3

Watch at 07:41

Open-source economics and the reality of renting capacity

GLM 5.2 is open source, whereas ChatGPT and Claude are closed models accessed by paying the provider directly. That said, GLM 5.2 is enormous at 753 billion parameters, so running it locally requires serious hardware and infrastructure that most people do not have. In practice, that means renting it online in a way that is not so different from paying Anthropic for Claude.

Even so, it is much cheaper than Claude. For a full day of heavy coding, the creator estimates it costs about one-fifth as much as Opus 4.8. Compared with Opus 4.8's pricing of $5 for input and $25 for output, GLM's cost profile is a large part of why people are reacting so strongly.

03 Slide 3 / 3

Watch at 15:07

A future where local models become standard

The creator predicts a future in which every company runs its own local models. He believes Anthropic has started to recognize that as well, which is why it is expanding its investment beyond the model itself into services and enterprise integration channels such as forward-deployed engineers. In his view, companies are beginning to realize that the model may not be the ultimate moat.

The performance gap between closed and open systems is still large, but he argues that it is closing quickly, and that watching the shift happen in real time is one of the most interesting parts of the current moment. He closes by saying he plans to cover more local models and open code options that are not tied to Claude Code's harness.

Editor's Take

Reports that GLM 5.2 is usable inside Claude Code's harness show value migrating from the model itself to the harness — the working environment around it. If heavy work runs at a fifth of Opus's cost, the realistic optimization is not staying on the expensive model but swapping models while keeping the environment. Anthropic's pivot toward services and integration support quietly confirms the same dynamic.

Source

Nate Herk | AI Automation

GLM 5.2 in Claude Code is Blowing My Mind

Published 6/19/2026 16 min 83,545 views

This article auto-summarizes the YouTube video's transcript with Claude. Please refer to the original video for nuance and exact wording.

Watch on YouTube

3 articles

Nate Herk | AI Automation

How Claude Is Creating a New Generation of Millionaires -- Without Writing a Line of Code

Vulcan, a company building software for government agencies, looks like the work of a hundred engineers -- but most of its team can't write a single line of code and built it all by describing what they wanted to Claude.

The Calum Johnson Show

Building a Membership Site With Claude Code in a Weekend, $20K in the First Week: 'You Don't Need the Terminal'

The guest says he built a membership platform, digitalcreatorclub.com, that would have previously cost $50,000-$100,000, in a single weekend using Claude Code -- and made about $20,000 in its first week.

tef

Claude Fable 5 vs. China's GLM 5.2: A Real Head-to-Head Building Three Games From Scratch

A head-to-head build-off between Claude Fable 5 and China's GLM 5.2, which some say now outperforms Fable 5.