AI has evolved over the past few months from a mere “chat assistant” into an “agent” capable of actually operating computers and completing complex tasks, and the extent of this evolution may exceed what either of us can imagine. Recently, a user shared on the X platform the astonishing process they witnessed firsthand of OpenAI’s GPT-5.5 using the Codex platform and the Computer Use function to instantly control a browser and complete an Amazon subscription cancellation, sparking heated discussion among netizens. This demo not only showcased the execution capabilities of AI agents in real-world scenarios but also drew renewed attention from the outside world to OpenAI’s latest model GPT-5.5 and its breakthrough progress in “computer use” capabilities.
GPT-5.5 Codex helped an online user successfully request a service cancellation and obtain a refund from a live customer service representative
According to user Chris (@chatgpt21) on X, he asked GPT-5.5 to help him cancel his Amazon subscription through the Codex platform. Throughout the process, the AI agent demonstrated impressive autonomous operation capabilities. Chris described sitting on his couch, watching GPT-5.5 automatically navigate to Amazon’s settings page, find the subscription management option, and execute the cancellation. However, things didn’t end after canceling the subscription. The AI agent further entered Amazon’s customer service page and proactively clicked on the “Chat with a customer service representative” option (the video below is sped up 10x).
I literally just watched GPT-5.5 via codex beat an Amazon customer associate in real time. 💀
I asked it to get me a refund, and I watched it navigate the settings, cancel the subscription, then it went step further into the help page.
I thought it was going to request a phone… pic.twitter.com/xepKx1HIxo
— Chris (@chatgpt21) April 29, 2026
When the Amazon customer service representative replied “Your subscription is currently active,” GPT-5.5 immediately responded with precision, explaining that the subscription shows as active because cancellation preserves access to the remaining billing cycle, but the user wanted to stop service immediately and receive a refund. The customer service rep ultimately confirmed the refund, and the AI agent ended the conversation without hesitation.
Chris wrote in a tweet: “This is the first time I’ve seen a human customer service representative get surpassed by AI in real-time.” He also added that the refund was $15 (approximately NT$495), which nearly offset the cost of the AI service within 5 minutes.
GPT-5.5’s Core Capabilities: Beyond Chatting
This test video gained widespread attention because it perfectly demonstrates the “Agentic Intelligence” that OpenAI emphasized when releasing GPT-5.5. According to OpenAI’s official introduction, GPT-5.5’s core upgrade directions include:
Computer Operation Skills (Computer Use)
GPT-5.5, paired with Codex’s computer control capabilities, can actually click, scroll, type text, take screenshots, and understand what’s on screen. This means AI is no longer just processing text messages—it can now operate web interfaces just like a human. In the Amazon cancellation case, GPT-5.5 used this capability to complete the entire process, from navigating to the settings page to chatting with customer service.
2. Tool Coordination and Multi-step Task Planning
OpenAI noted on their official blog that GPT-5.5 “can understand what you’re trying to do and take on more of the work itself.” It excels at writing and debugging code, online research, analyzing data, creating documents and spreadsheets, operating software, and completing tasks across tools. In the cancellation case, rather than just executing a single step, the AI autonomously planned a complete strategy of “cancel subscription → access customer service page → negotiate with a human agent → request a refund.”
3. Efficiency and Cost Advantages
According to benchmark data released by OpenAI, GPT-5.5 achieves 82.7% accuracy on Terminal-Bench 2.0 (a complex command-line workflow test), a significant improvement over the previous generation GPT-5.4’s 75.1%. Additionally, GPT-5.5 uses noticeably fewer tokens when completing the same Codex tasks, making it more efficient.
In Artificial Analysis’s Coding Index, GPT-5.5 delivers top-tier intelligence performance at half the cost of competing frontier models.
Safety and Ethical Considerations
OpenAI rolled out “the most powerful safety measures in history” alongside the release of GPT-5.5. According to official statements, GPT-5.5 underwent a full Safety and Preparedness Framework evaluation and collaborated with internal and external red team testers to conduct specialized assessments on advanced cybersecurity and biological capabilities.
In Amazon cancellation cases, AI agents act entirely under user instructions with clear goals—canceling subscriptions and seeking refunds—and involve no fraud or misconduct. However, the development of such capabilities has also raised concerns about potential misuse of AI agents. For instance, if AI can negotiate with real customer service representatives in real time and secure refunds, could it be weaponized for large-scale automated fraud or harassment in the future? (Or perhaps the future becomes a world where AI assistants haggle with AI customer service?)
Source: KOCPC Chinese