# **GPT-4.5: An Expensive Upgrade with Minimal Improvements**
## **Introduction**
OpenAI’s newest AI model, GPT-4.5, has been launched amidst considerable anticipation, yet initial assessments indicate that it may not deliver the groundbreaking advancements many expected. Though it is considerably pricier than its forerunner, GPT-4o, the enhancements in performance are only slight. Critics contend that GPT-4.5 highlights the dwindling returns associated with scaling conventional large language models (LLMs), prompting discussions about the future direction of AI innovation.
## **Performance vs. Cost: An Unsatisfactory Exchange**
GPT-4.5 stands as OpenAI’s most sophisticated traditional AI offering to date, but its performance boosts come with a hefty financial burden. The model incurs costs that are **30 times higher for input and 15 times greater for output** compared to GPT-4o. Based on OpenAI’s own assessments, GPT-4.5 shows little significant improvement over its predecessor across various tasks, causing some specialists to doubt the value of investing in larger models.
An unnamed AI professional remarked to *Ars Technica*, *”GPT-4.5 is a lemon!”*—a view shared by AI researcher Gary Marcus, who labeled the release a *”nothing burger.”* Even former OpenAI researcher Andrej Karpathy conceded that while GPT-4.5 surpasses GPT-4o, the advancements are nuanced and hard to measure.
## **Coding Performance: A Notable Deficiency**
A prominent drawback of GPT-4.5 lies in its lackluster performance in programming tasks. Independent evaluations by tech investor Paul Gauthier with the **Aider Polyglot Coding benchmark** revealed that GPT-4.5 placed **10th overall**, falling short of models like Claude 3.7 Sonnet and OpenAI’s own reasoning models, o1 and o3.
Moreover, GPT-4.5 has a **knowledge cutoff in October 2023**, which means it is unaware of the latest changes to programming languages and frameworks. This shortfall renders it a less dependable resource for software developers compared to alternative AI models.
## **GPT-4.5 vs. OpenAI’s Simulated Reasoning Models**
OpenAI’s benchmark data underscore GPT-4.5’s limitations in comparison to its **simulated reasoning models**, o1 and o3. For instance:
– **AIME Math Competition Scores:**
– GPT-4.5: **36.7%**
– o3-mini: **87.3%**
– **Cost Analysis (per million tokens):**
– GPT-4.5: **$75 (input), $150 (output)**
– o1 Pro: **$15 (input), $60 (output)**
– o3-mini: **$1.10 (input), $4.40 (output)**
These statistics indicate that OpenAI’s simulated reasoning models deliver **superior results at a significantly lower cost**, making GPT-4.5 an undesirable option for numerous use cases.
## **A Model Concentrated on “Vibes” Rather than Reasoning**
CEO of OpenAI **Sam Altman** sought to manage expectations by characterizing GPT-4.5 as a model that is *”high on vibes, low on reasoning.”* He admitted that while GPT-4.5 seems more engaging in conversational contexts, it does not perform well in analytical tasks or benchmark evaluations.
Altman also disclosed that OpenAI **lacked sufficient GPUs** to launch GPT-4.5 widely, further underscoring the model’s inefficacy.
## **The Conclusion of Traditional LLM Expansion?**
Given the underwhelming outcomes from GPT-4.5, OpenAI appears to be pivoting away from conventional LLMs. Altman has indicated that **GPT-5 will deviate from this trajectory**, opting instead to merge **non-reasoning LLMs with simulated reasoning models** such as o3.
This transition implies that OpenAI acknowledges the **diminishing returns** associated with merely enhancing model size and training datasets. The company is investigating **innovative architectures**, including inference-time reasoning and **diffusion-based AI models**, which could provide more effective and scalable alternatives.
## **Competition in the AI Landscape**
As GPT-4.5 struggles to validate its expense, rivals like **Anthropic’s Claude 3.7 Sonnet** are making strides. Claude 3.7 Sonnet has shown **greater efficacy** with a more optimized architecture, indicating that OpenAI may encounter increasing challenges from competing AI firms.
## **Conclusion: An Expensive Venture with Limited Prospects**
GPT-4.5 is now accessible to **ChatGPT Pro subscribers**, with plans to extend availability to **Plus, Team, Enterprise, and Education users** in the near future. However, OpenAI has not committed to sustaining the model in the long run, likely due to its high
