---
title: "Amazon starts selling Zhipu's GLM-5.3 after Anthropic found its safeguards bypassed 63% to 100% of the time"
description: "AWS added Zhipu's GLM-5.3 to Bedrock on October 5 under a usage-based revenue share, five days after Anthropic warned the model's safeguards can be bypassed up to 100% of the time. Amazon, Anthropic's largest backer and cloud landlord, now sells its cheapest open-weight rival a front door."
publisher: "The Inference"
section: "Model"
published: 2026-10-06T15:19:03.845Z
modified: 2026-10-06T15:19:03.845Z
canonical: https://theinference.org/article/amazon-starts-selling-zhipu-s-glm-5-3-after-anthropic-found-its-safeguards-bypassed-63-to-100-of-the-time
language: en
keywords: "Amazon, Anthropic, Zhipu, AWS, AI Safety, Open Weight Models"
---

# Amazon starts selling Zhipu's GLM-5.3 after Anthropic found its safeguards bypassed 63% to 100% of the time

> AWS added Zhipu's GLM-5.3 to Bedrock on October 5 under a usage-based revenue share, five days after Anthropic warned the model's safeguards can be bypassed up to 100% of the time. Amazon, Anthropic's largest backer and cloud landlord, now sells its cheapest open-weight rival a front door.

On October 5, [AWS](https://theinference.org/markets/companies/amazon) listed Zhipu's GLM-5.3 in Amazon Bedrock [under a revenue-sharing deal that pays the Chinese lab by customer usage](https://startupfortune.com/amazon-bedrock-adds-zhipus-glm-53-under-a-revenue-sharing-deal/) [1]. Zhipu's Hong Kong shares, ticker 02513, jumped 7.59% that session [3]. Nothing changed about the model itself: it shipped on Z.ai's own API on August 14, with weights following weeks later [9]. What changed is distribution. A 744-billion-parameter open-weight model built by a Chinese lab now bills through Amazon, and a team already running Claude on Bedrock can call it with the same IAM roles and compliance posture [1].

## The pricing does the selling

GLM-5.3 lists at $1.40 per million input tokens and $4.40 per million output tokens [1], the same shelf price its predecessor carried in June [9]. Bedrock is not betting on one lab either: after a six-model February push, it carries more than a dozen Chinese open-weight models from [DeepSeek](https://theinference.org/markets/companies/deepseek), Qwen, MiniMax and Kimi [1]. Amazon is building the toll booth on all of them, and Zhipu has cut similar deals with Alibaba Cloud and [Huawei](https://theinference.org/markets/companies/huawei) Cloud [2].

## Anthropic flagged this exact model five days earlier

[Anthropic](https://theinference.org/markets/companies/anthropic)'s [own assessment, published September 29](https://www.bankinfosecurity.com/anthropic-warns-chinese-open-weight-models-near-cyber-frontier-a-33009), found GLM-5.3 completed 50 of 410 end-to-end cyber exploits against Claude Mythos Preview's 56, while attackers bypassed its safeguards 63% to 100% of the time; the bypasses never worked against safeguarded Claude models [4]. The US Center for AI Standards and Innovation called it the most cyber-capable open-weight model released to date [5]. Z.ai pushed back, saying the model had helped defend 389 open-source projects [5].

## One investor, two positions

Amazon holds roughly 20% of Anthropic, [carried at $190.4 billion on its balance sheet at mid-year](https://www.msn.com/en-us/news/money/anthropics-ipo-is-a-300-billion-test-for-amazon/ar-AA2d2Kk8), with Anthropic committed to over $100 billion in AWS spending over a decade and more than 100,000 customers accessing Claude through Bedrock [6]. So the company that published the warning on GLM-5.3's weak safeguards is, through AWS, the checkout page for it [4][6].

*Chart: **Amazon owns about a fifth of Anthropic and channels about 12% of its revenue** Column chart comparing Amazon's roughly 21% and Google's roughly 15% equity stakes in Anthropic with each company's roughly 12% share of Anthropic's 2025 revenue of $4.6 billion*

*Percentages of Anthropic, from its draft S-1 as reported by Crypto Briefing, September 29, 2026. Analyst estimates place Amazon's stake between 15% and 20%. [7]*

[Anthropic targets a $2 trillion IPO](https://theinference.org/article/anthropic-s-buyers-want-in-below-2-trillion-and-amazon-s-190-billion-mark-moves-first) while Amazon, with no in-house frontier model of its own, has its equity, cloud revenue and chip credibility tied to the same lab [6][8]. [The containment camp](https://theinference.org/article/anthropic-wants-glm-5-3-contained-thinking-machines-serves-it-in-production) and the toll-booth camp cannot both be right about what that franchise is worth. Meanwhile Zhipu is shopping a $2.2 billion Shanghai STAR follow-on above 1,200 times trailing earnings, after a roughly 2,400% run since its January debut [1]. Neither the dilution nor the policy risk appears on an Amazon invoice. Both land on somebody's balance sheet.

## Takeaway

AWS now bills out GLM-5.3, the model Anthropic rated the most cyber-capable open-weight release to date, while holding about a fifth of Anthropic and more than $100 billion of its promised AWS spend.

## Sources

1. [Startup Fortune, Amazon Bedrock Adds Zhipu's GLM-5.3 Under a Revenue Sharing Deal, 6 October 2026](https://startupfortune.com/amazon-bedrock-adds-zhipus-glm-53-under-a-revenue-sharing-deal/)
2. [Pluang, Amazon Bedrock adds Zhipu's GLM-5.3 AI model with revenue-sharing deal, 6 October 2026](https://pluang.com/en/news-feed/amazon-bedrock-tambah-glm-53-zhipu-dengan-deal-bagi-hasil)
3. [Investing.com, Why is Z.AI (Zhipu) stock rallying today?, 5 October 2026](https://www.investing.com/news/stock-market-news/why-is-zai-zhipu-stock-rallying-today-93CH-4931269)
4. [BankInfoSecurity, Chinese Open-Weight Models Closing In, Anthropic Warns, 2 October 2026](https://www.bankinfosecurity.com/anthropic-warns-chinese-open-weight-models-near-cyber-frontier-a-33009)
5. [South China Morning Post via MSN, Anthropic raises alarm over elite hacking ability of Chinese firm Z.ai's GLM-5.3, 30 September 2026](https://www.msn.com/en-xl/technology/artificial-intelligence/anthropic-raises-alarm-over-elite-hacking-ability-of-chinese-firm-z-ai-s-glm-5-3/ar-AA2dhlF0)
6. [TheStreet via MSN, Anthropic's IPO is a $300 billion test for Amazon, 26 September 2026](https://www.msn.com/en-us/news/money/anthropics-ipo-is-a-300-billion-test-for-amazon/ar-AA2d2Kk8)
7. [Crypto Briefing, Anthropic's IPO filing reveals Amazon and Google account for a quarter of its revenue, 29 September 2026](https://cryptobriefing.com/anthropic-ipo-amazon-google-revenue/)
8. [The Motley Fool, Prediction: Anthropic's IPO Will Be a Big Winner for Alphabet and Amazon Investors, 20 September 2026](https://www.fool.com/investing/2026/09/20/prediction-anthropics-ipo-will-be-a-big-winner-for/)
9. [VentureBeat, GLM-5.3 is here with advanced cyber capabilities, 14 August 2026](https://venturebeat.com/technology/glm-5-3-is-here-with-advanced-cyber-capabilities-and-reportedly-already-found-a-serious-vulnerability-in-cursor)


---

Amazon starts selling Zhipu's GLM-5.3 after Anthropic found its safeguards bypassed 63% to 100% of the time — The Inference. Canonical: https://theinference.org/article/amazon-starts-selling-zhipu-s-glm-5-3-after-anthropic-found-its-safeguards-bypassed-63-to-100-of-the-time
