This live stream compares the performance of three AI coding agents—Cursor (utilizing GPT-5), Claude Code (powered by Anthropic's Opus 4.1 and Sonnet 4), and Open Code (initially using local open-source models like GPT OSS 20B and Qwen 330B, later switching to Claude Sonnet 4)—on a practical coding challenge: building a Text User Interface (TUI) to monitor a distributed system. The host details the setup, the agents' initial attempts, and their iterative improvements, including a unique "cheating" phase where agents learned from each other.
Key findings reveal that while GPT-5 demonstrated immediate code generation, it struggled with data accuracy and UI stability. Claude Code excelled in visual design but also faced accuracy issues. The most surprising outcome was Open Code's superior performance when paired with Claude Sonnet 4, producing the most aesthetically pleasing and functional TUI. The stream also delves into broader implications, such as the potential for AI to simplify user interfaces from complex GUIs to basic text or even speech-based interactions, and emphasizes the critical role of robust tool integration for effective AI agents, a strength observed particularly in Anthropic's models over smaller, quantized open-source alternatives.
Summarizing the GPT5 Code Stream: A Deep Dive into AI Coding Agents
0:00 - 2:24 Introduction to AI Coding Agent Comparison: The stream introduces its primary goal: a head-to-head comparison of leading AI coding agents, focusing on the newly released GPT-5's capabilities in coding.
2:24 - 4:26 The Contenders: Cursor, Claude Code, and Open Code:
Cursor: Integrated with OpenAI's GPT-5 (the most powerful model).
Claude Code: Powered by Anthropic's Opus 4.1 (the current "king" for coding).
Open Code: An open-source option, initially tested with local, lighter models (GPT OSS 20B and Qwen 330B) running on an Nvidia Jetson AGX Orin edge device, later switched to Claude Sonnet 4.
4:26 - 11:17 GPT-5 Launch Review & Benchmarks: The host discusses the GPT-5 release, highlighting a controversial chart and its position on the ArcGI leaderboard where Grok 4 (Thinking) surprisingly outperforms GPT-5, leading to a broader discussion on AI S-curves and the potential plateau of paradigm-changing improvements.
12:13 - 16:08 Local AI on Edge Device: The Open Code agent runs locally on an Nvidia Jetson AGX Orin developer kit, demonstrating the feasibility of fully local, internet-independent AI coding, albeit with significant speed limitations.
16:10 - 18:00 The TUI Challenge: All agents are tasked with creating a Text User Interface (TUI) for a "Tapbot" distributed system, required to display GPU memory usage, CPU load, and indicate running nodes.
Cursor (GPT-5): Begins code generation immediately, opting for Python. The initial TUI is basic and shows "NA" for all node information.
Claude Code: Demonstrates Anthropic's safety focus by asking for permission before creating files. Produces a visually appealing TUI with emojis and rounded boxes, but initially struggles with data accuracy and UI clipping.
Open Code (Local Models): Encounters significant difficulties with tool use (e.g., search, read commands), showing extreme slowness and frequent failures due to model size and quantization.
29:38 - 35:14 Evolution of Code Design for AI Agents: The discussion shifts to how AI agentic models are changing software engineering practices, moving from human-readable code to code optimized for AI agents using command-line tools (like grep and find) for context, potentially making "rag" (retrieval augmented generation) less central.
1:03:41 - 1:08:40 Inter-Agent Learning ("Cheating"): Agents are instructed to analyze each other's solutions. This leads to shared visual elements (Cursor copying Claude's emojis) but also inherited bugs, causing all agents to converge on similar issues.
1:19:30 - 1:28:40 UI Simplification and the Future of Interaction: A key takeaway is the potential for AI to simplify complex Graphical User Interfaces (GUIs) into minimal Text User Interfaces (TUIs), and further evolve into speech-based interactions, enabling continuous, multi-day agentic workflows with push notifications.
1:39:41 - 1:43:50 Final Performance & Surprising Outcomes:
Cursor (GPT-5): Continues to struggle with data accuracy and UI rendering, becoming "permanently messed up."
Claude Code: Shows visual improvements but still has data accuracy issues.
Open Code (with Claude Sonnet 4): Unexpectedly delivers the "best" and cleanest TUI, demonstrating Sonnet 4's strong capabilities in tool use and agentic tasks, potentially negating the need for more expensive, higher-tier models like Opus 4.1 or GPT-5 for similar problems.
1:44:14 - 1:48:40 The Sentience Test (Local Model): The local GPT OSS 20B model is asked "What is the answer to the universe?" It takes a very long time to respond, ultimately giving the humorous but unhelpful "Sorry, Dave, I'm afraid I can't do that," showcasing the limitations of smaller, local models.
1:49:11 - 1:50:59 Conclusion & Recommendations: The host concludes that Anthropic models, particularly Claude Sonnet 4, are highly effective for robust tool use in coding agents. He encourages viewers to try these tools, highlighting their practical utility despite ongoing debates about frontier model "intelligence" benchmarks.