This live-streamed video provides a hands-on comparative analysis of several AI coding agents, prompted by the recent release of OpenAI's GPT-5. The host evaluates three primary agents on the task of building a text-based user interface (TUI) for a project: Cursor (powered by GPT-5), Claude Code (using Anthropic's Opus 4.1), and the open-source Open Code agent. Initially, Open Code runs a small, quantized open-source model (GPT-OSS 20B) on a local Nvidia Jetson device to explore the feasibility of fully offline agents, but it struggles significantly with tool use.
The stream discusses the broader implications of the GPT-5 release, noting from the ARC-AGI benchmark that it doesn't represent a massive leap over competitors like Grok 4. A central theme is the evolution of software development UIs, shifting from complex graphical interfaces to simple, text-based agentic interactions. This is driven by models becoming proficient at using human-like terminal commands (grep, ls, find) instead of relying solely on RAG systems.
In the head-to-head comparison, both the GPT-5 and Opus 4.1 agents produce buggy or incorrect TUIs. Surprisingly, the best result comes from the Open Code agent after its model is switched to the less powerful Claude Sonnet 4, which creates a clean, functional interface. The host concludes that for practical agentic tasks, robust and reliable tool use—a demonstrated strength of Anthropic's models—is currently more critical than raw benchmark performance, and that the Cursor implementation of GPT-5 was disappointingly unstable.
Decoding GPT-5: A Live Comparison of AI Coding Agents
1:52 Introduction & The Contenders: The stream's goal is to test the new GPT-5 coding model by comparing three AI coding agents:
Cursor: A VS Code-based agent using GPT-5.
Claude Code: Anthropic's agent, set to use the powerful Opus 4.1 model.
Open Code: An open-source agent, initially tested with a local, open-source model (GPT-OSS 20B) running on an Nvidia Jetson.
4:31 GPT-5 Release & Benchmark Analysis: The host discusses the GPT-5 launch, including a "chart crime" in the presentation and the ARC-AGI leaderboard, where Grok 4 outperforms GPT-5 (albeit at a higher cost). This suggests the pace of progress may be slowing and OpenAI is no longer the undisputed leader.
11:17 Running an Agent Locally: A key experiment involves running the Open Code agent with a quantized model on an Nvidia Jetson edge device. This demonstrates the potential for fully local, offline AI agents that don't require an internet connection, though the model proves too slow and fragile for the task.
16:10 The Coding Challenge: All agents are given the same task: create a text-based user interface (TUI) for the "Tapbot" project. The TUI should display the status, CPU load, and GPU memory usage of various computer nodes in a distributed system.
21:34 Initial Agent Failures: The agents show different behaviors and failure modes.
The local open-source model (GPT-OSS) running on Open Code fails immediately, unable to correctly use the available software tools.
Claude Code asks for permission before creating files, reflecting its safety-oriented design.
Cursor (GPT-5) proceeds to generate code without asking for permission.
29:16 The Future of UIs is Text: The host discusses a key insight from the creator of Claude Code: the paradigm for human-computer interaction is shifting. Agents that use simple, human-like terminal commands (grep, ls) are proving more effective than those reliant on complex RAG indexing. This could lead to all software UIs becoming simplified from complex GUIs to simple text-box interfaces.
GPT-5 (Cursor): Generates a basic TUI, but all data fields are empty ("NA").
Opus 4.1 (Claude Code): Creates a much more visually appealing TUI with emojis and rounded boxes, but the layout is broken and clips the bottom of the table.
50:51 GPT-5 Provides Incorrect Information: After a round of feedback, the GPT-5 TUI now populates data, but the information is factually wrong. It reports incorrect CPU core counts, has the wrong VRAM amounts, and hallucinates GPUs on machines that don't have them.
1:05:04 An Agent Collaboration Experiment: The host prompts the agents to "cheat" by looking at each other's code. This tests if an agent can analyze another's solution to identify strengths and weaknesses to improve its own, a workflow similar to ensembling models.
1:18:42 The Surprising Winner: After the local models fail, the Open Code agent is switched to use the Claude Sonnet 4 model. This combination produces what the host deems the best TUI of the three—it is visually clean, functional, and superior to the results from the more powerful GPT-5 and Opus 4.1 models.
1:39:41 Conclusion - Tool Use Trumps Benchmarks: The stream concludes that for practical coding tasks, a model's ability to robustly and reliably use tools is more important than its raw intelligence score on benchmarks.
Anthropic's models (both Opus and Sonnet) excelled at this, proving very effective.
GPT-5 (via Cursor) was disappointing, proving unstable and buggy, with its terminal UI crashing and becoming unusable by the end of the test.
The real revolution highlighted is the shift to agentic, text-based interfaces, which will likely change how we interact with all software.