OpenAI's Brian Fioca on GPT-5 vs Codex, Evals & The Responses API |The Roo Cast- Nov 6, 2025 Ep. 29

In this episode of Roo Cast, host Hannes Rudolph is joined by RooCode's lead developer Matt and CEO Danny, along with special guest Brian Fiocca from OpenAI. Brian kicks things off by sharing his extensive background in the startup world (including YC and Rescue Time) and explains his current role on the Applied Startups team at OpenAI. He then reveals the surprising story of how he first discovered Roo Code: he was looking for open-source evaluation suites to test how models like GPT-Five perform inside different coding tools and found RooCode's eval harness.The conversation dives deep into a detailed comparison of GPT-5 and GPT-5 Codex, with Brian explaining the key architectural differences and why one is more adaptable while the other is hyper-optimized for a specific harness. The team then explores the new frontier of "Evals as a Service," discussing how to move beyond simple correctness benchmarks (like Sweebench) and create "performance review" style evals for agentic tasks. The discussion also covers the critical importance of the Responses API for preserving chain-of-thought, the future of context memory, and Brian's top recommendations for the RooCode codebase.Resources Mentioned: OpenAI: https://openai.com/ Y Combinator: https://www.ycombinator.com/ Zed (Editor): https://zed.dev/ Minimax: https://minimax.chat/ Anthropic: https://www.anthropic.com/ Vercel AI SDK: https://sdk.vercel.ai/ Slack: https://slack.com/ Linear: https://linear.app/