OpenAI releases the ‘largest and most knowledgable model’ GPT-4.5, with reduced hallucinations but high API prices

OpenAI has released the version of GPT-4.5 (19459073), a research preview for its latest and most powerful Large Language Model (LLM) designed for chat applications. Unfortunately, it’s far-and-away OpenAI’s Most expensive modelis the most expensive (more below).

It’s also not a “reasoning model,” or the new class of models offered by OpenAI, DeepSeek, Anthropic and many others that produce “chains-of-thought,” (CoT) or stream-of-consciousness-like text blocks in which they reflect on their own assumptions and conclusions to try and catch errors before serving up responses/outputs to users. It’s more of a classic LLM.

However, according to OpenAI CEO Sam Altman and co-founder Sam Altman’s Post on the social network GPT-4.5: “The first that feels like talking with a thoughtful person.” I have sat in my chair and gotten good advice from AI several times.

He cautioned, however, that the company was nearing the upper limit of its graphics processing units and had to restrict access as a consequence:

“The bad news: it is a huge, expensive model. We wanted to launch both the plus and pro versions at the same time but we are out of GPUs. We will add tens and thousands of GPUs to the system next week, and then roll it out for the plus tier. (Hundreds of thousands are coming soon, and we’re pretty sure you’ll use them all.) It’s not how we want to work, but it’s difficult to predict growth surges which lead to GPU shortages.

GPT-4.5 will be available to subscribers today. OpenAI’s most costly subscription tieris ChatGPT Pro ($200 USD/month) and developers can access it across all paid API levels. Next week, the plan is to expand to the less expensive Plus and Team ($20/$30 per month) tiers.

GPT 4.5 can access OpenAI’s ChatGPT Canvas Mode, and users are able to upload files and images. However, it does not have multimodal features such as voice mode, screensharing, or video.

Unsupervised learning advances AI

GPT 4.5 represents a significant step forward in AI, especially in unsupervised training, which improves the model’s abilities to recognize patterns, make connections, and generate creative insights.

OpenAI researchers said that during a livestream demo, the model was pre-trained in multiple data centers simultaneously, suggesting a similar decentralized approach to that of rival lab Nous Research. This training regime helped GPT-4.5 to learn how to produce more intuitive and natural interactions, follow the user’s intent more accurately, and demonstrate greater emotion intelligence.

This model builds on OpenAI’s previous work in AI scale, reinforcing that more data and compute power lead to better AI performance.

GPT-4.5 will produce fewer hallucinations than its predecessors (37.1% compared to 61.8% in GPT-4o), allowing it to be more reliable on a wide range of topics.

What makes GPT 4.5 different?

OpenAI claims that GPT-4.5 was designed to create warm and intuitive conversations. It is able to interact with users more effectively and has a better grasp of context and nuance.

Its expanded knowledge base, and improved ability of interpreting subtle cues, allows it to excel in a variety of applications, such as:

• Writing Assistance: Clarifying content, improving clarity, and generating new ideas.
• Programming assistance: Debugging and suggesting code improvements, as well as automating workflows.
• Problem solving: Giving detailed explanations and helping in practical decision making.

GPT 4.5 incorporates new alignment methods that improve its ability to understand user preferences and intent.

How to access GPT 4.5

ChatGPT users can select GPT 4.5 in the model selection on web, mobile, and desktop. OpenAI will start rolling out GPT-4.5 to Plus and Team users next week.

GPT-4.5 for developers is available via OpenAI’s API. This includes the assistants API and batch API. It supports features such as function calling, structured outputting, streaming, system message and image inputs. This makes it a versatile tool that can be used for a variety of AI-driven applications. Currently, it does not support multimodal features such as voice mode or video screen sharing.

Pricing and implications for enterprise leaders

Enterprises, and team leaders, stand to gain significantly from GPT-4.5’s capabilities. GPT-4.5’s lower hallucination rates, enhanced reliability, and natural conversational abilities can support a variety of business functions.

• Improved Customer Engagement: Business can integrate GPT 4.5 into support systems to achieve faster, more natural, and less error-prone interactions. Marketing teams can create high-quality content that is on-brand. AI automation can help with debugging, workflow optimization, and strategic decision making.
• Customization and scalability: API allows for tailored solutions, allowing enterprises to build AI driven solutions that are suited to their requirements.

The API allows for tailored implementations, enabling enterprises to build AI-driven solutions suited to their needs. Pricing for GPT-4ovia OpenAI’s API is shockingly high. It costs $75/$180 for a million input/output tokens, compared to just $2.50/$10 with GPT-4.5.

With other recent models, such as Anthropic’s Claude 3.7 and Google’s Gemini 2 Pro or OpenAI’s reasoning “o” series, (o1, o3 mini high, o3), it will be interesting to see if GPT 4.5’s value, especially when using the API, is worth its relatively high price.

Early reactions of fellow AI researchers are varied

GPT-4.5’s release has sparked mixed responses from AI researchers and tech lovers on the social networking site X. This is especially after a leaked version of its “system card”a technical document outlining the model’s training and evaluations, revealed a variety of benchmarks results before the official announcement.

OpenAI’s final system card, published after the leak, contains some notable differences. For example, the line “GPT-4.5 does not represent a frontier model but is OpenAI’s biggest LLM, which improves GPT-4’s computation efficiency by more than ten times” has been removed. The OpenAI’s official system card is available here, while a leaked version of the card is attached below.

Teknium, (@Teknium1), the pseudonymous founder of Nous Research, a rival AI model provider, expressed disappointment with the new model. He pointed out that the model showed minimal improvements when compared to the other leading LLMs in terms of measuring massive multitasking language understanding (MMLU).

He asked, “It’s now been over two years and 1,000s more capital has been invested since GPT-4… What happened?” Others pointed out that GPT-4.5 performed poorly in comparison to OpenAI’s O3-mini model for software engineering benchmarks. This raised questions about whether or not this release represents a significant improvement. Some users defended the model beyond its raw benchmarks.

Software developer Haider (@slow_developer),highlighted GPT-4.5’s 10x improvement in computational efficiency over GPT-4, and its stronger general purpose capabilities compared to OpenAI’s STEM-focused O-series models.

AI news poster Andrew Curran (@AndrewCurran_), took on a qualitative view and predicted that GPT-4.5 will set new standards in creative writing and thought, calling it OpenAI’s “Opus.” These discussions highlight a broader AI debate: Should progress be measured solely in benchmarks or do qualitative improvements such as reasoning, creativity, and human-like interaction hold greater value?

OpenAI continues to position GPT-4.5 in a research preview

so that it can gain deeper insights about its strengths and weaknesses. The company is committed to understanding the way users interact with this model and identifying unexpected uses.

According to OpenAI, “Scaling up unsupervised learning continues AI progress by improving accuracy, fluency, and reliability.”

The company is constantly refining its models and GPT-4.5 provides a solid foundation for future AI advances, especially in reasoning agents that use tools. OpenAI is actively evaluating GPT-4.5’s long-term role in its ecosystem, despite its impressive capabilities.

GPT-4.5 will offer significant improvements to users in various domains with its broader knowledge base and improved emotional intelligence. OpenAI is interested in how developers, businesses, and enterprises integrate the model with their workflows and apps.

With AI continuing to evolve, GPT 4.5 marks another milestone for OpenAI in its pursuit of more reliable, capable and user-aligned languages models, promising new innovation opportunities in the enterprise landscape.