OpenAI Launches Its Largest AI Model Yet in Research Preview

OpenAI has announced the launch of GPT-4.5, its largest AI model to date, code-named "Orion." The model, trained with more computing power and data than any previous OpenAI release, is available as a research preview to select users.

Despite its scale, GPT-4.5 does not qualify as a "frontier model," OpenAI stated in a white paper. The term is often used to describe AI systems pushing the boundaries of artificial intelligence capabilities.

Subscribers to OpenAI's ChatGPT Pro plan, which costs $200 per month, gained access to GPT-4.5 on Thursday. Developers on paid tiers of OpenAI's API can also start using the model immediately. Users subscribed to ChatGPT Plus and ChatGPT Team are expected to receive access next week, an OpenAI spokesperson said.

GPT-4.5 was built using the same unsupervised learning technique that OpenAI has employed for past iterations, from GPT-1 to GPT-4. Historically, increasing model size and computational resources has resulted in notable improvements in capabilities, particularly in mathematics, writing, and coding.

OpenAI claims GPT-4.5 has improved "world knowledge" and "higher emotional intelligence" compared to its predecessors. However, industry observers note that the performance gains from scaling AI models appear to be slowing. On multiple benchmarks, GPT-4.5 falls short of newer AI "reasoning" models developed by Chinese AI company DeepSeek, rival firm Anthropic, and OpenAI itself. Reasoning models are designed to simulate logical thinking, problem-solving, and decision-making processes, rather than simply generating responses based on statistical patterns, as traditional large language models do.

It's worth noting that the model is relatively costly to operate. OpenAI is charging developers $75 per million input tokens — roughly equivalent to 750,000 words — and $150 per million output tokens. In comparison, OpenAI's widely-used GPT-4o model costs significantly less at $2.50 per million input tokens and $10 per million output tokens. The company is assessing whether it will continue making GPT-4.5 available via its API in the long term.

OpenAI has positioned GPT-4.5 as an enhanced version of GPT-4o, but not a direct replacement. Although GPT-4.5 supports file and image uploads within ChatGPT, it does not yet integrate with OpenAI's two-way voice mode.

On OpenAI's SimpleQA benchmark, which tests factual accuracy, GPT-4.5 outperformed GPT-4o and some of OpenAI's own reasoning models. The company says the model is less prone to hallucinations—instances where AI generates incorrect or misleading information—than previous versions. However, some competing AI models still exceed its performance on specific benchmarks.

On programming-related tests, GPT-4.5 performed comparably to GPT-4o and OpenAI's o3-mini model in some cases, but lagged behind Anthropic's Claude 3.7 Sonnet and OpenAI's more advanced deep research models. It also failed to match the top AI models in academic reasoning benchmarks such as AIME and GPQA.

But OpenAI claims GPT-4.5 excels in areas that standardized benchmarks may not fully capture, such as understanding human intent and generating more natural, contextually appropriate responses.

The release of GPT-4.5 comes amid growing debate over the limits of traditional AI scaling. OpenAI co-founder and former chief scientist Ilya Sutskever previously suggested that AI models may have reached a point where simply increasing training data and computational power yields diminishing returns.

In response, AI developers — including OpenAI — have begun shifting focus toward reasoning models, which process information more deliberately rather than generating rapid responses. OpenAI plans to merge its GPT series with its reasoning-focused "o" models in a future release, likely starting with GPT-5 later this year.

While GPT-4.5 represents a significant step in AI model development, it may ultimately serve as a transitional product rather than a game-changing breakthrough. OpenAI CEO Sam Altman acknowledged the high costs and GPU constraints limiting GPT-4.5's rollout but reaffirmed the company's commitment to refining AI reasoning capabilities.

"We're sharing GPT-4.5 as a research preview to better understand its strengths and limitations. We're still exploring what it's capable of and are eager to see how people use it in ways we might not have expected," OpenAI said in a blog post.

About the Author

John K. Waters is the editor in chief of a number of Converge360.com sites, with a focus on high-end development, AI and future tech. He's been writing about cutting-edge technologies and culture of Silicon Valley for more than two decades, and he's written more than a dozen books. He also co-scripted the documentary film Silicon Valley: A 100 Year Renaissance, which aired on PBS.  He can be reached at [email protected].

Featured

  • The AI Show

    Register for Free to Attend the World's Greatest Show for All Things AI in EDU

    The AI Show @ ASU+GSV, held April 5–7, 2025, at the San Diego Convention Center, is a free event designed to help educators, students, and parents navigate AI's role in education. Featuring hands-on workshops, AI-powered networking, live demos from 125+ EdTech exhibitors, and keynote speakers like Colin Kaepernick and Stevie Van Zandt, the event offers practical insights into AI-driven teaching, learning, and career opportunities. Attendees will gain actionable strategies to integrate AI into classrooms while exploring innovations that promote equity, accessibility, and student success.

  •  black graduation cap with a glowing blue AI brain circuit symbol on top

    Report: AI Is a Must for Modern Learners

    A new report from VitalSource identifies a growing demand among learners for AI tools, declaring that "AI isn't just a nice-to-have; it's a must."

  • digital network with glowing blue and red lines, featuring multiple red arrows shifting in different directions

    Report: Attackers Change Tactics as Ransomware Payoffs Decline

    Attackers are changing tactics as they collect less money from ransomware payoffs, according to a new report from Chainalysis, a blockchain analytics firm.

  • digital textbooks displayed on multiple tablets

    Faculty Need Training, Time, and Tools to Make Course Content Accessible, Survey Finds

    In a recent survey by Anthology, only one in five faculty (22%) said they consistently consider accessibility when designing course materials. And just 11% felt they had the right tools and training to create accessible course content.