Hajra Naz
On Thursday, OpenAI unveiled GPT-4.5, the highly anticipated AI model code-named Orion. It is the largest model OpenAI has released, built using greater computing power and data than any previous iterations.
Despite its scale, OpenAI originally stated in its white paper that GPT-4.5 is not considered a frontier AI model.
Who Gets Access to GPT-4.5?
- ChatGPT Pro Subscribers – Immediate access for $200/month.
- Paid API Developers – Available starting today.
- ChatGPT Plus & ChatGPT Team Users – Rollout expected next week, according to OpenAI.
Read More: OpenAI’s GPT-5 Is Coming: Key Insights & Predictions
Scaling Up AI Training: A Breakthrough or Plateau?
The AI industry has been eagerly awaiting Orion, which serves as a test case for the effectiveness of traditional AI training techniques. GPT-4.5 follows the same foundational approach as its predecessors, using massive increases in computing power and data in the unsupervised pre-training phase.
Historically, this method has led to significant performance jumps across various domains such as mathematics, writing, and coding.
OpenAI claims that GPT-4.5 benefits from “deeper world knowledge” and “higher emotional intelligence.” However, some benchmarks suggest diminishing returns. Compared to other AI reasoning models from DeepSeek, Anthropic, and OpenAI itself, GPT-4.5 falls short in some areas.
Read More: Is DeepSeek better than ChatGpt?
Performance vs. Cost: Is GPT-4.5 Sustainable?
Running GPT-4.5 is costly, prompting OpenAI to evaluate its long-term viability within its API. Pricing details include:
- $75 per million input tokens (~750,000 words)
- $150 per million output tokens
By comparison, GPT-4o costs significantly less at $2.50 per million input tokens and $10 per million output tokens.
“We’re sharing GPT-4.5 as a research preview to better understand its strengths and limitations,” OpenAI stated in a blog post.
Mixed Performance Across Benchmarks
GPT-4.5 is not intended to replace GPT-4o, OpenAI’s widely used model. While it supports file and image uploads and ChatGPT’s canvas tool, it lacks features like ChatGPT’s two-way voice mode.
Strengths:
- Outperforms GPT-4o and OpenAI’s reasoning models (o1, o3-mini) on the SimpleQA benchmark for factual accuracy.
- Hallucinates less than most models, reducing misinformation risks.
- Excels in creative tasks, demonstrating a warmer, more natural tone in responses.
Weaknesses:
- Falls short against leading AI reasoning models such as Claude 3.7 Sonnet and DeepSeek’s R1 on academic tests like AIME and GPQA.
- Matches GPT-4o in coding challenges but struggles against OpenAI’s deep research model.
AI Reasoning Models: The Next Step in AI Evolution?
Experts suggest that traditional pre-training methods are approaching their limits. OpenAI co-founder Ilya Sutskever already expressed, “We’ve achieved peak data,” signaling the potential end of pre-training as we know it.
To address these challenges, OpenAI and other AI pioneers are centering on thinking models, which utilize expanded computing control to think through issues in a more organized way. These models take longer to process information but offer greater consistency and accuracy.
Looking Ahead: GPT-5 and Beyond
Later this year, OpenAI intends to combine its “o” reasoning series with its GPT series, beginning with GPT-5. Although GPT-4.5 was costly to develop, faced delays, and fell short of internal expectations, it serves as an essential stepping stone toward more advanced AI capabilities.
