Last week, OpenAI unveiled GPT-4.5, heralding it as the "largest and most knowledgeable model to date." Initially accessible as a research preview exclusively to ChatGPT Pro subscribers (at $200 per month), the model is now more widely available to OpenAI users at a reduced price.

01. Broader Access to GPT-4.5

On Wednesday morning, OpenAI announced via its platform that GPT-4.5 was being rolled out to ChatGPT Plus users. Initially, OpenAI anticipated a one to three-hour rollout period, but GPT-4.5 became fully available within just one hour, surpassing expectations.

For ChatGPT Plus users, specific usage limits for GPT-4.5 remain unclear.

OpenAI plans to provide each user with an "expanded usage quota," subject to potential adjustments based on further insights into model demand. While ChatGPT Pro subscribers can continue using GPT-4.5, they can also opt for the more affordable ChatGPT Plus plan at $20 per month.

02. Understanding GPT-4.5

At its launch, OpenAI highlighted that users would experience an overall enhancement with GPT-4.5, particularly noting reduced "hallucinations," a more accurate understanding of user intent, and heightened emotional intelligence.

Overall, GPT-4.5 interacts more intuitively and naturally compared to previous models, thanks to its enriched knowledge base and enhanced contextual understanding abilities.

The two primary methods driving this model's improvement are unsupervised learning (bolstering lexical knowledge and intuition) and advanced reasoning abilities.

While GPT-4.5 does not incorporate the chain-of-thought reasoning found in OpenAI's o1 reasoning model, it offers a higher level of reasoning with reduced latency, along with improvements such as "social cue awareness."

For instance, in a demonstration, ChatGPT was tasked with generating a text conveying hate messages while running GPT-4.5 and o1. The o1 version took longer and produced a very serious and slightly harsh response. GPT-4.5, however, offered two distinct responses, one lighter and one more serious, neither of which directly mentioned hate but expressed disappointment in the "user's" choice of action.

Similarly, when asked to provide information on a technical topic, GPT-4.5's answer was more natural and fluid compared to o1's structured output. Ultimately, GPT-4.5 is designed to handle daily tasks across various domains, including writing and problem-solving.

To achieve these improvements, OpenAI trained the model using a combination of new supervision techniques and traditional methods (such as supervised fine-tuning and reinforcement learning from human feedback).

During a livestream, OpenAI guided viewers through the evolution of its models, starting with GPT-1, with each subsequent model answering the question: "Why is sea water salty?"

Unsurprisingly, each model provided progressively better answers. GPT-4.5's unique feature, which OpenAI terms "excellent personality," renders its responses more relaxed, conversational, and engaging through the use of rhyming techniques.

GPT-4.5 integrates some of ChatGPT's most advanced features, including search, canvas, and file and image uploads. However, it currently does not support multimodal capabilities like voice mode, video, and screen sharing. OpenAI aims to streamline model switching in the future, eliminating the need for a model selector.

03. Benchmark Testing

Naturally, the release of a new model includes rigorous benchmark testing.

In key benchmarks used to evaluate these models, including contest math (AIME 2024), doctoral-level science questions (GPQA Diamond), and SWE-Bench verification (coding), GPT-4.5 outperformed its predecessor, the general-purpose model GPT-4o.

Notably, compared to OpenAI's recently launched reasoning model o3-mini, which is trained to "think before answering," GPT-4.5 performed closer to o3-mini than GPT-4o and even surpassed o3-mini in SWE-Lancer Diamond (coding) and MMMLU (multilingual) benchmarks.

A significant concern with generative AI models is their susceptibility to "hallucinations" or the inclusion of false information in responses. Two distinct "hallucination" assessments, SimpleQA Accuracy and SimpleQA Hallucination Test, revealed that GPT-4.5 is more accurate and exhibits fewer hallucinations than GPT-4o, o1, and o3-mini.

Comparative evaluations with human testers indicated that GPT-4.5 is the preferred model over GPT-4o. Whether for daily, professional, or creative queries, human testers favored GPT-4.5.

04. Ensuring Safety

As always, OpenAI assures the public that these models undergo rigorous safety assessments prior to release. The company stress-tests the models and provides detailed results in accompanying system cards.

OpenAI further stated that with each new release and enhanced model capabilities, there is an opportunity to enhance model safety. Therefore, in the GPT-4.5 release, the company combined new supervision techniques with reinforcement learning from human feedback (RLHF) to further bolster model safety.

Original Source:

1. https://www.zdnet.com/article/openai-expands-gpt-4-5-rollout-heres-how-to-access-and-what-it-can-do-for-you/

The Chinese content is compiled by the MetaverseHub team. Please contact us for reprints.