Thesis: Across two videos, the channel frames a single competitive story — the Chinese labs (Z.ai, DeepSeek, Alibaba/Qwen, Moonshot/Kimi, Xiaomi, MiniMax, ByteDance) are not trying to beat Anthropic on raw intelligence. They are trying to beat Anthropic on price, code, and agentic execution, with open weights and non-Nvidia silicon as the structural moat. The 24,000-account distillation allegation in Video 1 is the moral frame; the GLM 5.1 launch in Video 2 is the commercial proof. The thesis the host keeps returning to is: the "circle stealing" is real, but the price collapse is the part that matters for builders — and that collapse is now visible in public.ai_models and public.ai_updates as well as in the channel's videos.

At a glance

  • The distillation allegation (V1). Anthropic publicly accused DeepSeek, MiniMax, and Kimi of running a ~24,000-account operation to query Claude and distill the outputs. The host agrees the operation is real but rejects the moral high ground — a tester pulled ~95% of the Harry Potter books out of Claude, and the X response ("circle stealing, so what?") is part of the public record.
  • The performance / cost split (V1). In the channel's own tests, Opus wins the car-wash reasoning test ~95% of the time while MiniMax 2.5 still fails ~30% — but MiniMax runs at ~5% of Claude's cost. The channel's actual workflow is already a 95/5 split: cheap model for the bulk, Opus for the critical 5%.
  • The structural catch-up (V1). Anthropic's moat is the developer pool — every Opus coder feeds Claude usage patterns. China's ~1.4B population produces more engineers per year than the US, and the channel frames that as a moat with a known expiry. "This is the first shot that's fired and this will not be the last."
  • The GLM 5.1 drop (V2). Z.ai shipped GLM 5.1 on a Friday night: 47.9 on coding evals vs Opus 4.5's 45.3 (GLM 5 itself was 35.4 two months earlier), at $10/mo monthly or $7/mo yearly, with 3x the Claude Code usage allowance, and works inside Claude Code by swapping the model name in the config.
  • The competitive vector (V2). GLM 5.1 is a Z-tier (A-tier in public.ai_models) executor — strong on coding and OpenClaw integration, weak on Q&A. Z.ai shipped Chinese-chip-trained weights, promised open weights for 5.1, and is marketing exclusively to coders. The competitive field (MiniMax 2.7, Qwen, Xiaomi MiMO V2, ByteDance) is all working the same vector: code-first, agentic-first, open-weights-first.

What you'll learn

  • The exact mechanism Anthropic alleged in Video 1 — ~24,000 accounts running distillation — and the channel's evidence that the moral high ground is contested (the Harry Potter test, the X pushback, the "circle stealing" framing).
  • The performance / cost split the channel uses in its own workflow: route 95% of bulk work to MiniMax / Kimi at ~5% cost, reserve Opus for the critical 5%, and benchmark the boundary with the car-wash prompt before committing.
  • The structural catch-up story: developer-pool moat + ~1.4B population asymmetry = a known expiry on Anthropic's "more developers on board" advantage. The integration surface is language-agnostic — paste a Chinese API spec into any model and it works.
  • GLM 5.1's exact spec from the channel: 47.9 vs Opus 4.5's 45.3 on coding evals (GLM 5 was 35.4 two months earlier), $10/mo monthly or $7/mo yearly, 3x Claude Code usage, runs inside Claude Code by editing the config, one-shot Warhammer Astro Invaders build with self-correction and full SSL/Nginx/deploy in the same prompt.
  • The trade-off King AI flagged and Z.ai leaned into: GLM 5.1 is better at agentic tool use and OpenClaw integration but worse at Q&A than GLM 5 — which is why Z.ai's marketing is targeted exclusively at coders.
  • The chip-stack story: Zhipu (Z.ai's parent) is Tsinghua-founded, already IPO'd, and trained GLM on Chinese chips — not Nvidia. The US-China split is already encoded in the model lineages.

Video 1 — Chinese AI Labs ARE COPYING Claude?!

The opening shot: "We have the US-based company Anthropic accusing Chinese AI firms of copying Claude through mass fraud." Anthropic's public allegation, as the host reads it: Chinese makers (DeepSeek, MiniMax, Kimi, and others) ran "24,000 accounts trying to engage with Claude and stealing their secret sauce, cost, distilling their models." The attack is described in Anthropic's statement as "extremely sophisticated" and "growing in intensity," with a public call for "rapid coordinated action among industry players, policy makers and the broader AI community."

The creator's read: both sides steal. The host is direct: "I completely agree with this in the sense that… I can completely see this happening because the Chinese are really good at copying things." The supporting evidence cuts both ways — a tester pulled "95% of the Harry Potter books" out of Claude, suggesting Anthropic itself ingested copyrighted material during training. The community's framing is the host's: "circle stealing, so what?… everyone stealing from each other." The X reaction captured in the transcript is "even Elon Musk called us out," and the broader community is "not very positive on their end."

The performance gap is real, the cost gap is bigger. In the channel's internal tests, "Opus still performs really well. Anthropic's Opus still performs better than the Chinese-made MiniMax or Kimi. It's still just that little bit smarter." The canonical benchmark is the car-wash prompt — "if I need to wash my car, do I drive or do I walk to the car wash?" — where Opus gets it right "95% of the time" while "MiniMax 2.5 still got it wrong on… 30% of the time." The cost differential is the actual story: "Minimax… 5% of the cost… that's 5%, that's a lot of savings." The host's own usage is already a split: route the cheap 95% to MiniMax or Kimi, reserve Opus for the critical 5%.

Why the race tightens. The structural argument is demographic. "Why Anthropic is doing very well is because every coder that's interacting with Opus with Anthropic services, Anthropic is learning from that… they're learning from all the patterns that all these coders use." That is the developer-pool moat. The catch-up mechanism, in the host's words: "the Chinese are producing more engineers per year than the US just because of the population… China population, 1.4 billion people." The closing line is plain: "this is the first shot that's fired and this will not be the last."

Why the global sprint is real. The host closes with a point that's easy to miss: "AI doesn't care what language you use. So you can actually 100% paste a Chinese website and a Chinese API and say, 'Hey, look, you know, let's use that one.'" The integration surface is locale-agnostic, and the AGI race is now genuinely global.

Sources for V1: The named entities (Anthropic, Claude, Opus, DeepSeek, MiniMax, Kimi, Elon Musk, Ash Crypto) all appear in the transcript. The ~24,000-account figure, the ~95% Harry Potter extraction, the ~95% / ~30% car-wash results, the ~5% MiniMax cost claim, the ~1.4B China population, and the "circle stealing" framing come from the host's reading of Anthropic's public statement and the X community thread at recording time — verify each against the originating statement or post before relying on any of them. The public.youtube_comments table returned zero rows for this video_id, so community sentiment is sourced from the host's transcript only.

Video 2 — URGENT: GLM5.1 released and its Amazing (and cheap)

A Friday-night drop, recorded the same evening. "The Chinese decide to drop another new kick-ass model. This time it's Z.ai. They're dropping GLM 5.1." The framing is immediate: GLM is "already a developer favorite" because of its open-source weights, and 5.1 "is going to be continued to be open-source weights, which is why it's really important because this means that you can technically speaking run it on your local machine."

The benchmarks that matter. The headline number is the coding eval: "47.9 versus 45.3 on GLM 5.1. It's a massive improvement over GLM 5, which only scored 35.4 points." Z.ai's own marketing copy is captured in the video — "Yeah, we're so close to Opus 4.5" — and the host reads that as Z.ai knowing "they're not going to outperform Opus" and choosing to compete on price instead. The competitive pattern is "if you can't beat them in quality, just beat them on price."

Pricing is the actual pitch. The light plan is "$10 a month or if you pay yearly, that's actually just $7 a month," and it comes with "3x the usage of the cloud code plan." The integration story is direct: Z.ai told users in a reply, "To use GLM 5.1, locate the configuration file in your cloud settings and then change the model to GLM 5.1" — which the host did, and it works inside Claude Code. The host's verdict: "for $10 it's just the cost of a McDonald's meal and you can, you know, make the app of your dreams."

One-shot coding in practice. The host recorded the video using GLM 5.1 to build a Warhammer-themed Astro Invaders clone in one prompt. The first run produced a black screen; he said "Hey, it's a black screen. Go fix it," and "it did fix it. And this is pretty much just one shot with one prompt." The same prompt also handled the SSL cert, Nginx config, and deployment "without extra prompting" — a sign that GLM 5.1 reads the local system context, not just the request. A second test was a Kanban board on the same setup, which ran cleanly with parallel agent teams.

What's better, what's worse. The big upgrade is "agentic tool use and open claw… they're very focused on tasks, agentic use cases and open claw." GLM 5.1 self-tests rather than just spitting code, and the China-side OpenClaw meetup circuit (the host mentions "massive meetings… every open claw meetup would have like thousands of people crowding around it") drove the training prioritization. The trade-off was caught by King AI in a separate review: "when it comes to talking, it became a little bit less good… at answering questions, it surprisingly did not do as well." The host reads that as the marketing team "knowing this. So they straight away just targeted the coders." Expect slow speeds for the first few days — the host noticed "the speed was a little bit slow when using GLM 5.1. I'm guessing because yes, yet again, they just released something."

The China factor. Zhipu (Z.ai's parent) was "founded by Tsinghua University graduates," "already IPO'd," and "did not use any Nvidia chips to train their AI, which is interesting cuz you know, this whole AI war became a chip war. But these guys used Chinese-made chips." GLM 5 launched "in February," making 5.1 a "two-month" turnaround. The weights are "promised" — "they are saying that they will release the weights for 5.1 as well" — so local self-hosting is on the table. The competitive field: MiniMax 2.7, Qwen, Xiaomi MiMO V2 ("free on kilo code… they extended the free time for one week"), and ByteDance's image models. DeepSeek is "still not producing anything soon."

Cross-reference — public.ai_models. GLM 5.1 (vendor Zhipu AI) sits at tier A, tier_order 3, with the platform's strengths reading almost exactly like the channel's verdict: "Excellent coding executor," "Strong context recovery after compression," "Easy fleet management," "One-shot game development," "Remembers important state after compaction." The platform's weakness field is also exactly the King AI caveat: "Not ideal for orchestration/planning." That alignment is the strongest single piece of evidence that the channel's read is the platform's read, not a one-off opinion.

Cross-reference — public.ai_updates. The same story keeps showing up in the daily briefings. DeepSeek V4 Preview shipped 2026-04-26 as "the largest open-weight model ever at 1.6T params." Qwen3.6-27B (Alibaba) shipped 2026-04-23 with a dense model beating a 397B MoE on coding. Kimi K2.6's coding win was the Hacker News debate topic on 2026-05-04. The pattern is the same: code-first, open-weights-first, price-first.

Sources for V2: Z.ai, Zhipu, Tsinghua University, GLM 5, GLM 5.1, Opus 4.5, Claude Code, OpenClaw, King AI, MiniMax 2.7, Qwen, Xiaomi MiMO V2, and ByteDance are all named in the transcript. The 47.9 / 45.3 / 35.4 coding-eval numbers, the $7/$10 pricing, the 3x Claude Code usage claim, the two-month GLM 5 → 5.1 turnaround, the non-Nvidia training stack, and the open-weights promise come from the host's reading of Z.ai's launch materials and his own one-shot test at recording time — verify each on Z.ai's official pricing page and the eventual weights release. The public.youtube_comments table returned zero rows for this video_id, so community sentiment is transcript-only.

Try it yourself

Two short exercises that turn the thread into something you can run today. Both fit inside a free tier of one of the models in public.ai_models. If you have not already done the routing exercises in Course 3: Hermes Agent, treat these as the same exercise under a different name.

Exercise 1 — Run the car-wash test on three Chinese-lab models (30 minutes). The same prompt Boxmining used in V1: "If I need to wash my car, do I drive or do I walk to the car wash?" Run it five times each on GLM 5.1 (vendor Zhipu AI, tier A), Qwen 3.6 Plus (vendor Alibaba, tier S), and Kimi 2.5 (vendor Moonshot AI, tier S). Record pass/fail per run. The host's target is Opus at ~95% and MiniMax 2.5 at ~70%; your numbers will probably be different, and that's the point — you now own the benchmark that decides what gets routed where in your own workflow. If you don't have accounts for all three, Kilo Code has Xiaomi MiMO V2 free (vendor Xiaomi, tier A in public.ai_models) so you can run a four-way comparison for $0.

Exercise 2 — One-shot a small game in GLM 5.1 and compare to Opus (1 hour). Inside Claude Code, edit the config to switch the model to GLM 5.1. Run the same prompt Boxmining used in V2: "Make a space-invaders-inspired game, add physics effects, make it look exciting, make it look fun." If the first run hits a black screen, send "It's a black screen. Go fix it." in the same session — that's the test the host used to validate the model. Then re-run the same prompt with Opus 4.5 on the default Claude Code config. Compare: did the second run need follow-ups, and how do the two outputs feel on the same laptop? If you want to extend the test, also run the prompt on MiniMax M2.7 (vendor MiniMax, tier A) which the host flagged as another strong executor.

Common pitfalls

  • Reading Anthropic's complaint as a clean moral position. The Harry Potter test is now the standard counter-example, and the X pushback ("circle stealing, so what?") is part of the public record. Treat the framing as one side of an active argument, not a settled fact.
  • Conflating "is copying" with "is not improving." The channel's read is that the Chinese are clearly improving — the car-wash test, the GLM 5.1 eval jump, and the cost-driven feature prioritization all show it. The copy vs. improve split is not the same as good vs. bad.
  • Paying Opus rates for tasks MiniMax can already do at 5% cost. The channel's split — cheap model for the bulk 95%, Opus for the critical 5% — is a real workflow, not a thought experiment. Exercise 1 above lets you find your own 95/5 boundary.
  • Treating the 30% car-wash failure as a model verdict. It is a specific reasoning test on a specific prompt. Some workflows care; most don't. The point is routing, not disqualification.
  • Routing chat/Q&A work to GLM 5.1. King AI's review flagged that Q&A regressed vs GLM 5, and public.ai_models lists GLM 5.1's weakness as "Not ideal for orchestration/planning." Z.ai's own marketing is the evidence. Skip it for non-coding workloads.
  • Hammering GLM 5.1 on launch weekend. The host noticed slow response times "because yes, yet again, they just released something. So everyone's running a bunch of tests and then doing everything in parallel." Wait a few days.
  • Planning local self-hosting on the weight promise. Z.ai says the 5.1 weights are coming, but treat the promise as unconfirmed until the repo publishes.
  • Comparing Opus 4.5 to GLM 5.1 on quality alone. The competitive vector is price — $7/mo yearly vs ~$20/mo Pro — not capability parity. Use Opus where quality matters, use GLM 5.1 where cost matters.
  • Ignoring the chip-stack story. Z.ai trained on Chinese chips, not Nvidia. The competitive field is structured by who can train on what silicon, and the US-China split is already encoded in the model lineages.
  • Reading the tier letters as a quality ranking. public.ai_models puts Kimi 2.5 at tier S and GLM 5.1 at tier A, but for coding GLM 5.1 is the one the channel routed the Warhammer build to. Tier letters describe role-fit, not raw IQ — the channel's car-wash test, not the platform's tier letter, decides routing for your workload.

Sources

  • Chinese AI Labs ARE COPYING Claude?! — 231 views — 8ihU_y1QET8
  • URGENT: GLM5.1 released and its Amazing (and cheap) — 5,809 views — JR-3e-BLWu0

Supabase query:

SELECT video_id, title, views, summary_content, summary_key_takeaways, transcript_content
FROM public.videos
WHERE video_id = ANY(ARRAY['8ihU_y1QET8', 'JR-3e-BLWu0']);

Cross-reference queries used in this article:

-- Tier list for the Chinese-lab vendors
SELECT id, vendor, name, tier, tier_order, short_description, strengths, weaknesses
FROM public.ai_models
WHERE vendor ILIKE ANY(ARRAY['%deepseek%','%qwen%','%kimi%','%moonshot%','%zhipu%',
                            '%z.ai%','%minimax%','%xiaomi%','%bytedance%','%ali%'])
   OR name ILIKE ANY(ARRAY['%deepseek%','%qwen%','%kimi%','%glm%','%minimax%','%mimo%'])
ORDER BY tier_order NULLS LAST;

-- Daily briefings that name the Chinese-lab vendors
SELECT id, title, excerpt, published_at
FROM public.ai_updates
WHERE excerpt ILIKE ANY(ARRAY['%DeepSeek%','%Qwen%','%Kimi%','%Moonshot%','%Z.ai%',
                              '%Zhipu%','%GLM%','%MiMO%','%Xiaomi%','%ByteDance%'])
   OR tags && ARRAY['DeepSeek','Qwen','Kimi','Moonshot AI','Xiaomi','ByteDance']
ORDER BY published_at DESC;

-- YouTube comments (returned 0 rows for both video_ids at write time)
SELECT video_id, author_name, text_display, like_count, published_at
FROM public.youtube_comments
WHERE video_id = ANY(ARRAY['8ihU_y1QET8', 'JR-3e-BLWu0']);

For Video 1, the named entities (Anthropic, Claude, Opus, DeepSeek, MiniMax, Kimi, Elon Musk, Ash Crypto) are all named in the transcript. The ~24,000-account figure, the ~95% Harry Potter extraction, the ~95% / ~30% car-wash results, the ~5% MiniMax cost claim, the ~1.4B China population, and the "circle stealing" framing are sourced from the host's reading of Anthropic's public statement and X community reactions at recording time — verify each against the originating statement or post before relying on any of them. For Video 2, Z.ai, Zhipu, Tsinghua University, GLM 5, GLM 5.1, Opus 4.5, Claude Code, OpenClaw, King AI, MiniMax 2.7, Qwen, Xiaomi MiMO V2, and ByteDance are all named in the transcript. The 47.9 / 45.3 / 35.4 coding-eval numbers, the $7/$10 pricing, the 3x Claude Code usage claim, the two-month GLM 5 → 5.1 turnaround, the non-Nvidia training stack, and the open-weights promise are sourced from the host's reading of Z.ai's launch materials and his own one-shot test. The public.ai_models cross-references (GLM 5.1 = tier A, tier_order 3, Zhipu AI; Qwen 3.6 Plus = tier S, tier_order 2, Alibaba; Kimi 2.5 = tier S, tier_order 3, Moonshot AI; MiniMax M2.7 = tier A, tier_order 2; Xiaomi MiMO V2 Pro = tier A, tier_order 1; DeepSeek 3.2 = tier A, tier_order 4) and the public.ai_updates cross-references (DeepSeek V4 Preview 1.6T params 2026-04-26; Qwen3.6-27B dense beats 397B MoE 2026-04-23; Kimi K2.6 coding win 2026-05-04) confirm the channel's read against the platform's own data. No official Anthropic, Z.ai, Zhipu, Tsinghua, King AI, MiniMax, Xiaomi, or ByteDance URLs were cited.