This Day in AI Podcast We Built Microsoft Teams in 23 Minutes (And You Can Use It) & GPT 5.4 Impressions - EP99.37
142 snips
Mar 6, 2026 They test GPT-5.4âs new capabilities, context windows, and surprising tradeoffs in agentic speed. They describe rebuilding full SaaS apps from single prompts, including a working video-chat team app and a Trello clone. They debate steep Pro pricing, compare models like Opus and Gemini, and predict a wave of cheap agentic tools that could upend many SaaS businesses.
AI Snips
Chapters
Transcript
Episode notes
GPT 5.4 Expands Context And Reasoning
- GPTâ5.4 adds a ~1,000,000 token context and tuned reasoning levels, enabling much larger, designâfriendly knowledge work outputs.
- Michael notes 128k out and multiple reasoning tiers (lowâextra high) that trade speed for deeper problem solving.
Speed Up Agents By Choosing Lower Reasoning
- Lower reasoning level to speed up agentic loops when you need throughput rather than deep thinking.
- Chris runs five agent tabs and drops 5.4 to its lowest thinking level to get faster, still-useful results.
Reduce Token Cost By Using Tool Search
- Use tool search (only short tool descriptors up front) to save tokens in intense agent loops and fetch full tool definitions only when selected.
- Chris implemented this pattern and OpenAI added an official Tool Search API to support it.