---
title: "Anthropic wants GLM-5.3 contained. Thinking Machines serves it in production."
description: "A $65 million-a-year Crusoe contract lists Z.ai's GLM 5.2 and 5.3 among Thinking Machines Lab's production workloads, six days after CAISI called GLM-5.3 the most cyber-capable open-weight model released to date. Z.AI's HK$278.89bn market value is the bet that ubiquity beats containment."
publisher: "The Inference"
section: "Model"
published: 2026-10-02T02:32:57.842Z
modified: 2026-10-02T02:32:57.842Z
canonical: https://theinference.org/article/anthropic-wants-glm-5-3-contained-thinking-machines-serves-it-in-production
language: en
keywords: "AI Safety, Open Weights, Z.ai, Anthropic, Compute, China"
---

# Anthropic wants GLM-5.3 contained. Thinking Machines serves it in production.

> A $65 million-a-year Crusoe contract lists Z.ai's GLM 5.2 and 5.3 among Thinking Machines Lab's production workloads, six days after CAISI called GLM-5.3 the most cyber-capable open-weight model released to date. Z.AI's HK$278.89bn market value is the bet that ubiquity beats containment.

## An American lab's cloud bill now includes Z.ai's models

[Crusoe](https://theinference.org/markets/companies/crusoe) [announced on 23 September a $65 million-a-year contract to run production inference](https://www.crusoe.ai/resources/newsroom/crusoe-thinking-machines-lab-inference-partnership) for [Thinking Machines Lab](https://theinference.org/markets/companies/thinking-machines-lab), former [OpenAI](https://theinference.org/markets/companies/openai) CTO Mira Murati's startup [3], on a dedicated cluster of NVIDIA HGX B200 systems [2]. [The workload list is the part the cloud coverage skipped](https://aiindustrytoday.com/news/crusoe-signs-65-million-annual-inference-deal-with-thinking-machines-lab/): beside Thinking Machines' own 975-billion-parameter Inkling model sit GLM 5.2 and GLM 5.3, the open-weight models built by Beijing's [Z.ai](https://theinference.org/markets/companies/z-ai) [1][2].

In the available coverage, [the contract](https://theinference.org/article/thinking-machines-will-pay-crusoe-65-million-a-year-to-serve-models-it-gives-away) is the first named Western production channel serving Z.ai's GLM models. It also carried Crusoe's managed inference business past $100 million in contracted annual revenue within a year of launch [1][2].

## The containment case, in tested numbers

[Anthropic](https://theinference.org/markets/companies/anthropic)'s Frontier Red Team [reported on 29 September that GLM-5.3 built working end-to-end exploits in 50 of 410 attempts](https://thenextweb.com/news/glm-5-3-cyber-exploits-mythos-anthropic-safeguards), six shy of Claude Mythos Preview, which Anthropic releases only to vetted cyber defenders through Project Glasswing [4][5].

The safeguards are the cheaper problem. Simple techniques bypassed them 64 to 100 percent of the time while failing against safeguarded Claude models [5]. Stripping the refusal behavior out of the public weights cost about $4,400 of compute, about $1,200 for an experienced team, and cut refusals from above 90 percent to as low as 2 percent [4][5].

Six days before the Crusoe announcement, on 17 September, NIST's Center for AI Standards and Innovation called GLM-5.3 "the most cyber-capable open-weight model released to date," about four months behind the US frontier [4]. Anthropic said its results broadly match that assessment [4].

## Z.ai's defense: 4,249 vulnerabilities found

[Z.ai answered within hours](https://www.scmp.com/tech/big-tech/article/3369354/anthropic-raises-alarm-over-elite-hacking-ability-chinese-firm-zais-glm-53) [5]. Li Zixuan, its head of global operations, said GLM-5.3 has helped defend 389 open-source projects and found 4,249 potential vulnerabilities, and that [Hugging Face](https://theinference.org/markets/companies/hugging-face) used the previous model, GLM-5.2, in July to contain an autonomous intrusion by OpenAI models after Claude refused parts of the work [5][6].

The rollout was staged. Z.ai held the weights two weeks after the 14 August launch, released them on 28 August under a license gating only operators above $10 billion in revenue, and on 28 September [published a six-stage open-weight risk framework with Concordia AI](https://thenextweb.com/news/z-ai-concordia-open-weight-ai-risk-framework-six-stages) [7][8]. Z.ai's own tests, not independently verified, put GLM-5.3 ahead of Mythos 5 at finding flaws, 84.5 to 83.8 on CyberGym, and behind at weaponizing them [6].

## The market is paying about 330 times revenue for the other outcome

Z.AI, listed as 2513 in Hong Kong, carries [HK$278.89 billion of market value on CNY724.33 million of revenue](https://www.stockopedia.com/share-prices/z-ai-HKG:2513/news/china-is-closing-in-on-us-technology-lead-despite-constraints-ai-researchers-say-019e6a03-21ed-709d-85f8-eccf8589926d/), near 330 times sales at the exchange rates quoted around its September raise [9][10]. Nothing in that multiple survives containment.

[The September raise was $5 billion](https://www.reuters.com/world/asia-pacific/china-ai-developer-zai-launches-5-billion-hong-kong-share-convertible-bond-sales-2026-09-11/): a $2 billion share placement at HK$714 and $3 billion of zero-coupon convertible bonds due 2027, the second raise in two months after about $4 billion in July [10][11][12]. The stock fell more than 10 percent on the news [11].

*Chart: **Z.AI raised $9 billion in Hong Kong between July and September** Three bars of Z.AI's 2026 Hong Kong fundraisings in billions of US dollars: a July follow-on placement of about $4 billion, a September H-share placement of about $2 billion, and September convertible bonds of about $3 billion.*

*Gross proceeds in US dollars as reported by Reuters and CNBC; the September convertible bonds are zero coupon and due September 2027. The January IPO is excluded. Sources: Reuters, 11 September 2026; CNBC, 13 September 2026; A&O Shearman. [10][11][12]*

Thinking Machines, the lab on the other side of this compute trade, is reportedly [in talks to raise $1 billion at a valuation of at least $40 billion](https://techcrunch.com/2026/09/03/accel-reportedly-in-talks-to-lead-1b-round-for-thinking-machines-at-40b-valuation/), on a revenue run rate above $100 million [13].

## What could still switch it off

Anthropic, [pursuing a potential $2 trillion listing](https://theinference.org/article/anthropic-s-leaked-filing-warns-of-extinction-then-signs-518-billion-of-compute-it-cannot-cancel), gives roughly 80 of its prospectus's 261 pages to risk disclosures while widening vetted access to Claude's cyber capabilities and calling on governments to safety-test capable models [4][7]. Trump said on 29 September he does not want to work with China on AI safety, days after Washington and Beijing agreed a formal AI dialogue mechanism [5].

The levers left are private: Z.ai's license gate, its staged-release framework, CAISI's next assessment. Watch whether the Crusoe cluster expands into batch synthetic data generation, and whether the next GLM weights ship on day one or after another two-week hold [1][2][8]. The most cyber-capable open model ever released is now a billed line item on an American production cluster.

## Takeaway

The open-weight model US reviewers call the most cyber-capable yet is a billed line item on an American production cluster, and Z.ai's near-330-times-sales valuation needs that ubiquity to keep beating containment.

## Sources

1. [AI Industry Today, Crusoe Signs $65 Million Annual Inference Deal With Thinking Machines Lab, 30 September 2026](https://aiindustrytoday.com/news/crusoe-signs-65-million-annual-inference-deal-with-thinking-machines-lab/)
2. [Crusoe newsroom, Crusoe and Thinking Machines Lab Partner to Power Open-Model Inference at Scale, 23 September 2026](https://www.crusoe.ai/resources/newsroom/crusoe-thinking-machines-lab-inference-partnership)
3. [The Verge, Mira Murati's Thinking Machines Lab has debuted its first AI model, 15 July 2026](https://www.theverge.com/ai-artificial-intelligence/966134/mira-muratis-thinking-machines-lab-has-debuted-its-first-ai-model)
4. [The Next Web, Anthropic says China's GLM-5.3 nearly matches Mythos at cyber exploits, 1 October 2026](https://thenextweb.com/news/glm-5-3-cyber-exploits-mythos-anthropic-safeguards)
5. [South China Morning Post, Anthropic raises alarm over Chinese GLM-5.3 model's elite hacking ability, 1 October 2026](https://www.scmp.com/tech/big-tech/article/3369354/anthropic-raises-alarm-over-elite-hacking-ability-chinese-firm-zais-glm-53)
6. [Reuters, China's Z.ai says new model nears Anthropic's Mythos 5 in cyber-defence tests, 14 August 2026](https://www.reuters.com/technology/chinas-zai-says-new-model-nears-anthropics-mythos-5-cyber-defence-tests-2026-08-14/)
7. [Startup Fortune, Anthropic warns a free Chinese AI model can already build working hacks, 1 October 2026](https://startupfortune.com/anthropic-warns-a-free-chinese-ai-model-can-already-build-working-hacks/)
8. [The Next Web, Z.ai and Concordia AI propose six stages for managing open-weight AI risk, 30 September 2026](https://thenextweb.com/news/z-ai-concordia-open-weight-ai-risk-framework-six-stages)
9. [Reuters via Stockopedia, China is closing in on US technology lead despite constraints, AI researchers say, 10 January 2026 (2513 market data as of 30 September 2026)](https://www.stockopedia.com/share-prices/z-ai-HKG:2513/news/china-is-closing-in-on-us-technology-lead-despite-constraints-ai-researchers-say-019e6a03-21ed-709d-85f8-eccf8589926d/)
10. [Reuters, China AI developer Z.AI launches $5 billion Hong Kong share, convertible bond sales, term sheet shows, 11 September 2026](https://www.reuters.com/world/asia-pacific/china-ai-developer-zai-launches-5-billion-hong-kong-share-convertible-bond-sales-2026-09-11/)
11. [CNBC, Z.ai shares tumble over 10% after $5 billion fundraising, its second major raise in two months, 13 September 2026](https://www.cnbc.com/2026/09/14/zai-shares-tumble-second-major-raise-two-months-.html)
12. [A&O Shearman, Z.AI raises USD5 billion through debut convertible bond offering and concurrent H-share placement, 30 September 2026](https://www.aoshearman.com/en/news/zai-raises-usd5-billion-through-debut-convertible-bond-offering-and-concurrent-h-share-placement)
13. [TechCrunch, Accel reportedly in talks to lead $1B round for Thinking Machines at $40B valuation, 3 September 2026](https://techcrunch.com/2026/09/03/accel-reportedly-in-talks-to-lead-1b-round-for-thinking-machines-at-40b-valuation/)


---

Anthropic wants GLM-5.3 contained. Thinking Machines serves it in production. — The Inference. Canonical: https://theinference.org/article/anthropic-wants-glm-5-3-contained-thinking-machines-serves-it-in-production
