Generation AI
GPT-5 review: what is it good for, reasoning on by default, hallucinations down, prompt rules change
Episode Summary
In this episode, Generation AI hosts Ardis Kadiu and JC Bonilla, joined by Petar Djordjevic, break down OpenAI's biggest release in two years - GPT5. They explore how this isn't just one model but four different versions (GPT5, Pro, Mini, and Nano) unified through an intelligent router system that automatically selects the right model for each task. The team discusses the dramatic 45% reduction in hallucinations compared to GPT4, the competitive pricing that undercuts rivals, and what this means for higher education institutions using AI. They share firsthand experiences implementing GPT5 at element451, explain why developers need to completely rethink their prompting strategies, and reveal how OpenAI is positioning itself against Claude in the coding assistant market. This episode matters because GPT5 fundamentally changes how institutions can deploy AI - making advanced reasoning capabilities accessible to all users while dramatically reducing costs and errors.
Episode Notes
GPT5 Launch and Model Architecture (00:00:00)
- OpenAI announces GPT5 after multiple delays
- Four model variants: GPT5, GPT5 Pro, GPT5 Mini, and GPT5 Nano
- Introduction of intelligent router system that automatically selects models
- Launch issues with router sending queries to wrong models initially
The Router Revolution: No More Model Selection (00:05:06)
- How the router uses previous ChatGPT usage signals to train selection
- Product decision to remove model dropdown confusion for users
- Small model in front makes decisions based on task complexity
- Users can influence selection by asking model to "think hard"
Dramatic Improvements in Accuracy (00:12:43)
- 45% hallucination rate vs GPT4's 80% rate
- Better data quality and reinforcement learning improvements
- Focus on agentic behavior and context gathering
- Tool calling accuracy improvements for real-world applications
Three Key Enhancement Areas (00:24:34)
- Coding: Direct competition with Claude and Anthropic's models
- Writing: Shorter, more concise, better quality outputs
- Medical/Healthcare: Improved analysis of health documents and test results
- Each area received specialized reinforcement learning
Developer Implementation Challenges (00:18:13)
- Markdown disabled by default requiring explicit instructions
- Shorter instructions work better than detailed prompts
- Need to rethink system prompts and instruction patterns
- Different behavior requires rewriting existing implementations
Pricing and Competitive Positioning (00:29:13)
- GPT5 offers best price-to-performance ratio in market
- 1/12 the cost of competing models like Claude Opus 4.1
- Free tier users get access to GPT5 with routing
- Pro tier ($200/month) provides research-grade intelligence
Real-World Implementation at element451 (00:20:11)
- Immediate deployment for summarization and classification tasks
- Evaluation ongoing for higher-stakes applications
- Benefits of pluggable AI architecture for new models
- Different models for different latency requirements
Market Impact and User Adoption (00:40:03)
- Free user reasoning model usage jumped from 1% to 7% in days
- Plus users increased from 7% to 24% reasoning model usage
- Traffic doubled overnight causing serving challenges
- OpenAI deprecating all previous models to focus resources
The Future of AI Assistants (00:43:22)
- Evolution from assistant to "chief of staff" capability
- Model knows when to act and how hard to think
- Implications for higher education automation
- Why institutions should adopt GPT5 immediately
- - - -
Connect With Our Co-Hosts:
Ardis Kadiu
https://www.linkedin.com/in/ardis/
https://x.com/ardis
Dr. JC Bonilla
https://www.linkedin.com/in/jcbonilla/
https://x.com/jbonillx
About The Enrollify Podcast Network:
Generation AI is a part of the . If you like this podcast, chances are you’ll like other Enrollify shows too!
Enrollify is made possible by Element451 — The AI Workforce Platform for Higher Ed.
Learn more at element451.com.