# Build Hour: GPT-5 Source: https://www.youtube.com/watch?v=ITMouQ_EuXI Recap page: https://rapidrecap.app/video/ITMouQ_EuXI Generated: 2025-09-03T01:01:08.873+00:00 --- ## Quick Overview GPT-5 significantly enhances coding capabilities, agentic task completion, and user steerability, offering a step-function increase in code quality, especially for front-end UI development. Key features include a "minimal" reasoning parameter for latency, plain text tool calls, and a "verbosity" parameter to control output detail. The new Responses API is crucial for leveraging GPT-5's full potential, enabling better performance in long-running tasks by maintaining conversational context through "chain of thought" reasoning tokens, which also improves caching. Charlie Labs demonstrated GPT-5's power by building a functional Minecraft clone and improving their coding agent, Charlie, with a 29% improvement in PR creation over GPT-3.5. **Key Points:** - GPT-5 represents a "step function" increase in code quality, particularly for front-end UI development, and is described as a "true coding collaborator" with enhanced steerability. - The new Responses API is highlighted as essential for unlocking GPT-5's capabilities, offering features like "minimal" reasoning for reduced latency and plain text tool calls, moving beyond the older Completions API. - GPT-5 excels at long-running agentic tasks by maintaining a coherent "train of thought" through chain of thought reasoning tokens, which are passed back to the model, preventing "amnesia" and improving performance on complex tasks. - The "verbosity" parameter allows control over the detail and readability of the model's output, with "verbosity high" producing more readable code and better error handling, beneficial for coding use cases. - Charlie Labs showcased GPT-5's capabilities by building a functional Minecraft clone using CodeX and demonstrated significant improvements in their autonomous coding agent, Charlie, achieving up to a 29% increase in PR creation performance compared to GPT-3.5. - Prompting with GPT-5 requires precision, as it interprets instructions very literally; avoiding conflicting instructions and using specific formatting like XML can lead to better results, with metaprompting aiding in refining prompts by asking the model for explanations of its behavior. - Future developments aim for models that can work in a loop to iterate on apps, better understand task complexity for longer or shorter horizons, and improved agentic capabilities for tasks spanning hours, days, or even weeks. **Context:** This build hour session introduces and demonstrates the capabilities of GPT-5, OpenAI's latest model. The session features speakers from OpenAI, including Christine from the startup marketing team, Bill, and Eric from the posting research team, who showcase GPT-5's advancements. The event also includes a guest appearance by Riley, founder of Charlie Labs, who shares how his company leverages GPT-5 for their autonomous coding agent. The primary goal is to empower developers and founders with best practices and tools to scale their companies using OpenAI's APIs and models, with a strong emphasis on coding and agentic tasks. ## Detailed Analysis The GPT-5 build hour highlights GPT-5 as a significant advancement in AI models, particularly for coding and agentic tasks. It's described as a "true coding collaborator" with a "step function" increase in code quality, excelling at front-end UI development and offering enhanced steerability. New features include a "minimal" reasoning parameter for improved latency and plain text tool calls, eliminating the need for JSON escaping. The new Responses API is presented as the optimal way to interact with GPT-5, offering a stateful or stateless approach and utilizing "chain of thought" reasoning tokens to maintain context in long-running tasks, preventing model "amnesia" and boosting performance. The "verbosity" parameter controls output detail and readability, with "verbosity high" being beneficial for code. Demos included building a Minecraft clone with CodeX, showcasing GPT-5's ability to generate complex applications from scratch. Charlie Labs demonstrated their autonomous coding agent, "Charlie," which uses GPT-5 to interact with platforms like GitHub, Linear, and Slack, improving their PR creation by up to 29% over GPT-3.5. Prompting advice emphasizes precision, avoiding conflicting instructions, and using metaprompting to refine model behavior. The session also touched on future aspirations for models that can iterate on applications in loops and handle tasks over longer time horizons with increased agentic capabilities. ### Introduction to GPT-5 - Discusses new capabilities including coding, agentic tasks, and new features like the 'minimal' reasoning parameter and plain text custom tools. ### GPT-5 Capabilities and Benchmarks - Highlights GPT-5 as a coding collaborator with improved code quality, front-end UI capabilities, and steerability, mentioning benchmarks like Sweetbench and Ador Polyglot. ### Agentic Tasks and Reasoning - Explains GPT-5's proficiency in long-running agentic tasks with tool calls, its instruction following, and self-correction, along with the "minimal" reasoning parameter for balancing intelligence and latency. ### Responses API vs. Completions API - Contrasts the new Responses API with the older Completions API, emphasizing the former's richer features, improved developer experience, and stateful nature for maintaining context via "chain of thought" tokens. ### Chain of Thought and Caching - Details how "chain of thought" tokens enable models to remember intermediate reasoning steps, crucial for long tasks, and improve prompt caching performance, making the system cheaper, more intelligent, and faster. ### Verbosity Parameter - Explains the "verbosity" parameter's role in steering the final output and tool calls, with "verbosity high" producing more readable code and better error handling. ### CodeX Demo - Showcases GPT-5 generating a functional Minecraft clone from scratch using CodeX CLI, demonstrating its ability to plan, execute, and iterate on code based on user prompts. ### Prompting Best Practices - Offers tips for GPT-5 prompting, including precision, avoiding conflicting instructions, using XML for structure, metaprompting for refinement, and allowing room for planning and self-reflection, with an emphasis on controlling "eagerness" and "persistence". Prompting Tips: Use precise instructions, leverage "medium" reasoning effort initially, consider XML structure, use metaprompting to ask 'why', provide planning steps, and control "eagerness/persistence". Controlling Eagerness: Discusses how to manage model "eagerness" or "persistence" through prompt design to either complete tasks end-to-end or involve the user more collaboratively, similar to pair programming. Tool Preamble: Explains the use of tool preambles to prompt GPT-5 for explanations of its tool calls, enhancing transparency and allowing for better monitoring and control of agentic workflows, especially with custom tools like MCP servers mentioned. Instruction Following Example: Illustrates prompt optimization by fixing conflicting instructions in a hypothetical patient scenario, showcasing how metaprompting can clarify ambiguities and improve model adherence to user intent. Charlie Labs Demo: Riley, founder of Charlie Labs, demonstrates their TypeScript-focused coding agent, "Charlie," which integrates with GitHub, Linear, and Slack, leveraging GPT-5 for autonomous task completion, including bug detection, issue creation, and PR generation, showing significant performance gains over previous models and competitors like Claude Opus 41. Performance Metrics: Charlie Labs reports a 16% improvement from GPT-3.5 to GPT-5 for PR creation, reaching up to 29% after optimization, and a 5% improvement in PR review evaluations over GPT-3.5, while Claude Opus 41 performed 46% worse in their tests. Q&A and Future Outlook: Addresses questions on context engineering for coding, emphasizing summarization techniques for long contexts, and discusses future directions for AI models, including continuous iteration on apps, improved agentic capabilities, and understanding task complexity over different time horizons.