What Happened
OpenAI released GPT-4.5, described as its largest and most compute-intensive pre-training run to date. The model focused on improving the breadth and depth of world knowledge, reducing hallucinations, and enhancing the model's ability to understand nuanced or ambiguous prompts. It was initially available to ChatGPT Pro subscribers and via API.
Why It Matters
GPT-4.5 represented OpenAI's push toward improving the "base intelligence" of large models through massive-scale unsupervised pre-training, as a complement to the reasoning-focused o-series models. OpenAI positioned it as having the best "EQ" (emotional intelligence) and factual grounding of any of its models, suggesting a dual-track strategy: scaling pre-training for knowledge and world modeling, while scaling test-time compute for reasoning.
Technical Details
- Architecture: Not fully disclosed; continuation of GPT-4 architecture with significant scaling
- Focus areas:
- Reduced hallucination rates compared to GPT-4 and GPT-4 Turbo
- Improved performance on knowledge-intensive tasks
- Better calibration and uncertainty awareness
- Enhanced instruction following and nuance understanding
- Context window: 128K tokens
- Availability: ChatGPT Pro tier, then broader API access
- Positioning: Described by OpenAI as the last major "non-reasoning" model before the shift to o-series and beyond