OpenAI Chief Research Officer Mark Chen: GPT 4.5 is Live and Scaling Isn’t Dead

Episode Summary

Executive Summary: OpenAI’s Mark Chen framed GPT-4.5 as the next step in OpenAI’s “predictable scaling” path: a larger, more capable model that improves knowledge work, creativity, and emotional intelligence while complementing—not replacing—reasoning models. The conversation emphasized dual-track progress (scaled models plus reasoning), the role of efficiency and architecture, and OpenAI’s belief that frontier models will continue to power better products and agents.

Main Topics: GPT-4.5 as a scaling milestone (Priority: 5/5): Chen positioned GPT-4.5 as the latest proof that OpenAI can keep scaling unsupervised learning, describing it as an order-of-magnitude step up from prior models. Why it is not GPT-5 (Priority: 5/5): OpenAI says GPT-4.5 reflects the current state of scaling in unsupervised learning, while GPT-5 will likely combine that with advances in reasoning and other capabilities. Dual-track strategy: scaling plus reasoning (Priority: 5/5): Chen argued OpenAI now has two complementary axes of progress: bigger pretrained models and reasoning models, with each supporting the other. Use cases: knowledge work, writing, coding, science (Priority: 4/5): GPT-4.5 was described as especially strong in productivity, knowledge work, creative writing, some coding tasks, and certain scientific domains. Efficiency and model architecture (Priority: 4/5): The discussion touched on serving costs, inference optimization, and architectural techniques like mixture of experts as key to making large models practical. Benchmarking and emotional intelligence (Priority: 4/5): Beyond standard benchmarks, OpenAI highlighted new qualitative strengths such as emotional intelligence, better conversation style, and improved ASCII art and writing. Frontier models vs niche/smaller models (Priority: 3/5): Chen defended OpenAI’s focus on frontier models while also acknowledging a portfolio approach that includes smaller, cheaper models for many use cases. OpenAI talent and internal bench strength (Priority: 2/5): Chen addressed reports of departures by saying OpenAI still has a world-class research organization and strong internal talent bench.

Key Arguments: GPT-4.5 continues OpenAI’s predictable scaling paradigm and represents the latest step after GPT-3, 3.5, and 4. OpenAI does not view scaling and reasoning as opposing approaches; reasoning benefits from broad knowledge, so the two paradigms are complementary. The model has not hit a scaling wall; OpenAI says it is still seeing expected returns from added compute, data, and optimization. GPT-4.5 is better suited than reasoning models for some tasks requiring immediate responses, strong world knowledge, creative writing, and certain coding/scientific workflows. Efficiency improvements matter separately from core capability, and OpenAI is always pushing the inference stack to reduce serving costs. Mixture-of-experts and other architectural improvements are applicable to both frontier foundation models and reasoning models. OpenAI believes frontier intelligence still matters because the last increments of capability can materially improve agentic systems and deep research products. Smaller and mini models remain important for cost-sensitive use cases, so the market is not purely about giant general-purpose models. The company is not just optimizing benchmarks; it is also exploring softer qualities like emotional intelligence and user resonance to understand real-world utility. OpenAI’s internal research culture remains strong despite talent departures, and the company sees movement in the AI field as natural.

Data Points: GPT-4.5 release timing: Today for Pro users; next week for Plus, Team, Enterprise, and EDU - OpenAI rollout schedule mentioned at the end of the interview Performance jump: “Order of magnitude” improvement - Chen described GPT-4.5 as similar in jump size to GPT-3.5 to GPT-4 User preference for productivity/knowledge work: 60% preference - Chen said people preferred GPT-4.5 over GPT-4.0 by this margin for productivity and knowledge work User preference for some use cases: Almost 70% preference rate - Chen cited stronger preference for GPT-4.5 versus GPT-4.0 in certain comparisons Model size: OpenAI’s largest model yet - Transcript repeatedly notes GPT-4.5 as the largest model OpenAI has released Time horizon for use-case discovery: 1–2 months - Chen said some use cases, such as creative writing, will need testing over the next one or two months Company history reference: 4.5 years - Host noted this was the show’s first OpenAI interview after 4.5 years Reasoning duration: Several minutes - Chen contrasted reasoning models with immediate-response models like GPT-4.5 Human benchmark comparison: 99.9 percentile to best in the world - Chen used math as an example of the value of frontier capability

Pivotal Quotes: "GPT 4.5, really, it signifies the latest milestone in our predictable scaling paradigm." — Mark Chen: Opening explanation of what GPT-4.5 represents "We now have two different axes on which we can scale." — Mark Chen: Chen explaining why GPT-5 may combine unsupervised learning and reasoning "Our mission isn’t just about pushing the biggest, most costly models. It’s about having that and also a portfolio of models that people can use cheaply for their use cases." — Mark Chen: Discussion of frontier models versus smaller niche models

Implications: OpenAI is betting that future AI progress will come from both larger foundation models and reasoning systems, with frontier capability driving better products like agents and deep research while smaller models handle cost-sensitive workflows.