Hey there, Seb here.
OpenAI just dropped GPT-4.5, their “largest and best model for chat yet” – and it’s a fascinating shift in direction.
Here’s the breakdown.
The Big Picture: Two Paths to AI Intelligence
OpenAI is now explicitly pursuing two complementary approaches to advancing AI:
- Reasoning (O1 series): Teaching models to think step-by-step before responding
- Unsupervised Learning (GPT series): Building deeper world knowledge and intuition
GPT-4.5 represents a massive scaling up of the unsupervised learning approach, resulting in a model that doesn’t explicitly “reason” like O1 but is “inherently smarter” with broader knowledge and understanding.
The Hard Numbers: Performance Breakthrough
The results are striking:
- Factual Accuracy: 62.5% on SimpleQA (vs. 38.2% for GPT-4o)
- Hallucination Rate: Down to 37.1% (vs. 61.8% for GPT-4o)
- Human Preference: Won 63.2% of professional queries and 57% of everyday queries against GPT-4o
- Creative Tasks: Won 56.8% of creative intelligence evaluations
What’s fascinating is that GPT-4.5 still lags behind the reasoning-focused O1 models on complex STEM problems, but excels at tasks requiring broad knowledge, creativity, and human-like intuition.
The User Experience: “EQ” and Natural Conversation
The most noticeable difference might be how it feels to talk with GPT-4.5:
- More Natural Conversations: Responses flow more like human dialogue
- Higher “EQ”: Better at reading emotional context and responding appropriately
- Better Intent Understanding: Grasps what you’re asking for, even when imperfectly expressed
- More Concise: Gives shorter, more focused answers without unnecessary detail
This was beautifully demonstrated in their side-by-side comparison where someone said they were “going through a tough time after failing a test.” GPT-4o provided a comprehensive, bulletpoint-heavy response with advice, while GPT-4.5 gave a warm, brief, empathetic reply that invited further conversation.
Practical Strengths: What It Does Best
OpenAI highlights these as GPT-4.5’s particular strengths:
- Improved Writing: Enhanced creativity and aesthetic intuition
- Programming: Better at understanding code context and user intent
- Practical Problem Solving: More intuitive grasp of real-world situations
- Creative Collaboration: Natural brainstorming and idea generation
- “Agentic Planning”: Multi-step tasks and complex workflows
The Technical Achievement: Engineering at Scale
Making GPT-4.5 required solving massive engineering challenges:
- Multi-datacenter training: Too big for a single location
- Advanced optimization: Including aggressive low-precision techniques
- New alignment methods: Using data from smaller models to teach larger ones
- Custom inference systems: To keep responses fast despite the size
Availability Timeline
- Available now: ChatGPT Pro users and API developers on paid tiers
- Next week: Plus and Team users
- Following week: Enterprise and Edu users
For API developers, OpenAI notes that GPT-4.5 is significantly more expensive to run than GPT-4o, and they’re “evaluating whether to continue serving it long-term” based on usage patterns and feedback.
Feature Support
GPT-4.5 works with:
- Search (for up-to-date information)
- File and image uploads
- Canvas (for writing and code)
- Function calling
- Structured outputs
- Vision capabilities
It does not currently support:
- Voice Mode
- Video
- Screensharing
What This Means for the Future
OpenAI believes these two approaches—reasoning and unsupervised learning—will eventually merge. As they put it: “As models like GPT-4.5 become smarter and more knowledgeable through pre-training, they will serve as an even stronger foundation for reasoning and tool-using agents.”
The most telling line from the announcement was this: “With every new order of magnitude of compute comes novel capabilities.” OpenAI seems genuinely unsure of GPT-4.5’s full potential, stating they’re “eager to see how people use it in ways we might not have expected.”
The Takeaway: Why This Matters
GPT-4.5 represents a significant philosophical shift in AI development. Instead of just teaching models to “think harder,” OpenAI is showing that massive scaling of unsupervised learning creates its own form of intelligence—one that’s less methodical but perhaps more intuitive and human-like.
For everyday users, this means more natural, helpful interactions. For developers, it offers a foundation for building applications that feel warmer and more emotionally intelligent.
The future likely involves hybrid systems combining both approaches: the intuitive, knowledge-rich foundation of models like GPT-4.5 with the careful reasoning capabilities of the O1 series.
We’re entering an era where AI doesn’t just calculate better—it understands better.
Stay ahead of the curve,
Seb
P.S. What will you build with GPT-4.5? Hit reply and let me know your ideas—I might feature the most innovative ones in our next issue!
