This live stream video pits the newly released GPT-5 against its main competitors in a head-to-head coding challenge. The host compares three AI coding agents: Cursor (powered by GPT-5), Claude Code (powered by Anthropic's Opus 4.1), and the open-source Open Code agent running various models, including a small open-source model (GPT-O OSS 20B) on a local NVIDIA Jetson edge device. The challenge is to create a functional and well-designed Text-based User Interface (TUI) for a robotics project.
Throughout the stream, the agents' different approaches, successes, and failures are observed. The local open-source models consistently struggle with tool use, highlighting their current limitations for complex agentic tasks. While both GPT-5 and Claude Opus 4.1 make progress, they also encounter significant bugs and eventually get stuck in a shared local minimum after being prompted to learn from each other's code. Surprisingly, the best result is produced by the Open Code agent when paired with the less powerful Claude Sonnet 4 model. The stream concludes that robust tool-use training, a strength of Anthropic's models, is currently more critical for agent performance than raw model intelligence, and that the future of user interfaces is trending towards this simplified, agent-driven, text-based paradigm.
Comparing AI Coding Agents: A GPT-5 Deep Dive
01:53 The Challenge Setup: The stream's goal is to test the coding capabilities of the newly released GPT-5. The host sets up a three-way comparison between coding agents:
Cursor: using OpenAI's new GPT-5 model.
Claude Code: using Anthropic's top-tier Opus 4.1 model.
Open Code: an open-source agent, initially tested with a local GPT-O OSS 20B model running on an NVIDIA Jetson AGX.
04:11 GPT-5 Release Discussion: Before coding, the host discusses the GPT-5 launch, noting that on some benchmarks like ARC-AGI, it is slightly outperformed by xAI's Grok 4. This suggests that OpenAI is no longer the undisputed leader and that progress at the frontier of AI is hitting the top of the "S-curve," where major leaps become more difficult.
16:10 The Coding Task: The agents are tasked with creating a Text-based User Interface (TUI) for a robotics project called "Tapbot." The TUI must display the status, CPU load, and GPU memory usage for several different computer nodes in the system.
22:40 Initial Agent Failures: The local open-source model running on Open Code fails immediately, unable to correctly use the available software tools. The host theorizes that these smaller models are overfit to a specific set of tools and lack the robustness of larger, proprietary models.
29:16 The Power of Human-like Commands: The host discusses a key insight from the creator of Claude Code: modern agents work best not by indexing a codebase (a technique called RAG), but by using human-like terminal commands such as grep, find, and ls to explore the code and gather context dynamically.
41:14 First Iteration Results: Both Cursor (GPT-5) and Claude Code (Opus 4.1) produce initial versions of the TUI.
GPT-5's TUI is functional but basic, with placeholder data.
Claude's TUI is aesthetically superior, using rounded boxes and emojis, but has layout issues where the text gets cut off.
59:11 A Collaborative Experiment: After several rounds of feedback where both agents struggle with fetching correct data, the host instructs them to analyze each other's code. This "ensembling" approach is designed to let them learn from each other's strengths and weaknesses, but it backfires as they both copy the same flawed logic, leading them to the same error.
1:17:10 The Unexpected Winner: After the main competitors get stuck, the host runs the task on the Open Code agent, this time using the mid-tier Claude Sonnet 4 model. Surprisingly, this combination produces the best TUI of the stream—a clean, functional, and aesthetically pleasing interface with emojis that outperforms both GPT-5 and the more powerful Claude Opus 4.1.
1:39:33 Final Conclusion: Tool Use Over Raw Intelligence: The host concludes that Claude remains the "winner." The experiment shows that a model's ability to robustly use tools is more critical for agentic coding tasks than its raw intelligence or benchmark scores. Anthropic's models, even the less powerful Sonnet 4, appear to have superior post-training for this specific, practical skill. GPT-5, in this test, did not demonstrate a clear advantage.