How Does GPT-4.5 Compare to GPT-4o? Key Differences and Improvements 196

小编 ·

How Does GPT-4.5 Compare to GPT-4o? Key Differences and Improvements

In the early hours Beijing time, OpenAI held a roughly 14-minute livestream to unveil its latest large model, GPT-4.5. In its press release, OpenAI described it as the largest and best chat model the company has ever produced, with over 10x the compute efficiency of GPT-4. Let's take a look at the specific changes and improvements compared to GPT-4o.

How Does GPT-4.5 Compare to GPT-4o? Key Differences and Improvements

At 4:00 AM Beijing time, OpenAI held a livestream event of about 14 minutes to release its latest large model GPT-4.5. In the press release, OpenAI stated that this is the largest and best chat model in the company's history, with over 10 times the computational efficiency improvement compared to GPT-4. Let's take a look at the specific changes and improvements compared to GPT-4o.

GPT-4.5 Release.webp

What are the changes between GPT-4.5 and GPT-4o?

  • Knowledge accuracy and hallucination rate: In the AIME2024 benchmark, GPT-4.5 scored 36.7%, while GPT-4o only scored 9.3%, showing a significant improvement in knowledge accuracy. For questions like "What was humanity's first language," GPT-4.5 will honestly say it's an unsolved mystery, with fewer hallucinated answers.

  • World knowledge and understanding: Through various optimizations, GPT-4.5 has broader knowledge and deeper understanding capabilities, no longer talking nonsense, with improved reliability.

  • Human preference and conversation ability: GPT-4.5 is more popular in human preference tests. It can better understand human thoughts, converse with higher emotional intelligence, grasp subtle cues and implied expectations more delicately, and has stronger aesthetic intuition and creativity.

  • Technical advancements: GPT-4.5 has new scalable technologies that can train powerful models using smaller model data, with improvements in steerability, detail understanding, and natural conversation ability.

  • Reasoning and application capabilities: Although GPT-4.5 is not a specialized reasoning model, it can answer simple yet challenging knowledge questions quite well, and handles code writing and physics problem-solving decently, though it falls short of dedicated reasoning models like Claude 3.7 in some areas.

  • Price: Using GPT-4.5 costs much more than GPT-4o. For input, GPT-4.5 charges $75 per million tokens versus $3 for GPT-4o; for output, GPT-4.5 charges $150 per million tokens versus $15 for GPT-4o. GPT-4.5 costs several times more than GPT-4o.

  • Features and limitations: GPT-4.5 has more features, such as search, file and image uploads, canvas writing, and coding capabilities. However, it does not support multimodal functions like voice, video, and screen sharing, making it inferior to GPT-4o in this regard.

Below is a comparison table between GPT-4.5 and GPT-4o:

FeatureGPT-4.5GPT-4o
Release DateFebruary 27, 2025March 15, 2023
Model ScaleTransformer MoE architecture with 1T activated parametersRelatively small parameter scale
Training DataFull internet content up to June 2024Relatively limited training data
Computational EfficiencyImproved by over 10 timesLower computational efficiency
Hallucination Rate37.1% in the SimpleQA benchmarkRelatively high hallucination rate
Knowledge BreadthExpanded knowledge reserves through unsupervised learningRelatively limited knowledge reserves
Emotional IntelligenceCapable of recognizing user emotions and providing warmer responsesLower emotional intelligence
Conversation NaturalnessMore natural conversation experience, understanding implied meaningRelatively mechanical conversation experience
CreativityExcellent performance in writing, design, etc.Weaker creativity
Multimodal CapabilitiesSupports file and image uploads, handles writing and coding tasksRelatively weaker multimodal capabilities
Price$75 per million input tokens, $150 per million output tokens$2.50 per million input tokens, $10 per million output tokens

The Development History of OpenAI's GPT Series Models

GPT-1:

  • Release Date: 2018

  • Features: Could barely generate coherent text, marking the beginning of the GPT series.

GPT-2:

  • Release Date: 2019

  • Features: Limited capability, sometimes seemed chaotic, but began to show some ability to generate coherent text.

GPT-3:

  • Release Date: 2020

  • Features: A significant leap, with greatly improved text generation capabilities, becoming an important milestone in the GPT series.

GPT-3.5 Turbo:

  • Release Date: 2022

  • Features: Provided more accurate and concise answers on certain questions, though still with room for improvement.

GPT-4:

  • Release Date: March 2023

  • Features: An overall better feel, with more creative word choices, improved prompt understanding, and a slight reduction in hallucination frequency. Although the improvements are subtle, overall performance is better.

GPT-4.5:

  • Release Date: February 2025

  • Features: A more human-like conversational experience, with improved emotional understanding and emotional intelligence, significantly enhanced knowledge reserves and capabilities. Although the improvements are subtle, they are still exciting.