OpenAI launches GPT-4.5 'Orion,' its largest AI model yet

OpenAI has announced the launch of GPT-4.5, the highly anticipated AI model code-named Orion, marking the company's most expansive model to date. Trained using more computing power and data than any of its predecessors, GPT-4.5 represents a significant step in OpenAI's ongoing AI research.

Despite its scale, OpenAI initially stated in a white paper that it does not classify GPT-4.5 as a frontier model. However, hours after the model's release, the company removed the line from the white paper, leaving its official stance on the model's classification unclear.

Subscribers to ChatGPT Pro, OpenAI's $200-a-month plan, now have access to GPT-4.5 as part of a research preview. Developers on paid tiers of OpenAI's API can also begin using the model immediately. Broader access is expected next week for ChatGPT Plus and ChatGPT Team users.

GPT-4.5 was developed using the same pre-training technique, dramatically increasing data and computing power during unsupervised learning, that powered earlier models like GPT-4, GPT-3, and GPT-2. Historically, this scaling approach has delivered notable improvements in AI capabilities, from writing and mathematics to coding.

However, OpenAI acknowledges that the returns from scaling may be tapering off. While GPT-4.5 showcases a 'deeper world knowledge' and 'higher emotional intelligence,' it lags behind newer AI reasoning models from companies like DeepSeek and Anthropic on certain AI benchmarks.

Performance and cost

GPT-4.5 has shown mixed performance across key AI tests. On OpenAI's SimpleQA benchmark , which measures accuracy on factual questions, the model outperformed GPT-4o and OpenAI's o1 and o3-mini reasoning models. However, it fell short when compared to Perplexity's Deep Research model.

In coding tests, GPT-4.5 matched GPT-4o and o3-mini on the SWE-Bench Verified benchmark but did not surpass OpenAI's deep research model or Anthropic's Claude 3.7 Sonnet. On the SWE-Lancer benchmark, which evaluates AI's ability to develop software features, GPT-4.5 outperformed GPT-4o and o3-mini but again fell behind deep research.

GPT-4.5's performance on academic tests like AIME and GPQA also placed it below leading AI reasoning models such as o3-mini, DeepSeek's R1, and Claude 3.7 Sonnet. Nevertheless, the model either matched or outperformed non-reasoning models in math and science-related tasks.

OpenAI highlights qualitative improvements, noting that GPT-4.5 exhibits stronger emotional intelligence and creative capabilities. The model responded in a 'warmer and more natural tone' to human prompts and excelled in creative tasks like writing and design. In an informal test, GPT-4.5 successfully generated an SVG image of a unicorn, a task GPT-4o and o3-mini struggled with. It also delivered the most socially sensitive response to a prompt about coping with failure.

High costs and future plans

Despite its strengths, GPT-4.5 is costly to run. OpenAI is charging developers $75 per million input tokens (roughly 750,000 words) and $150 per million output tokens, significantly higher than GPT-4o's $2.50 and $10 rates, respectively. The company is evaluating whether to continue offering GPT-4.5's API long term.

GPT-4.5 supports several advanced features, including file and image uploads and ChatGPT's canvas tool. However, it currently lacks ChatGPT's realistic two-way voice mode.

Looking ahead, OpenAI plans to merge its GPT and 'o' model series, with GPT-5 expected later this year. While GPT-4.5 may not lead AI benchmarks outright, OpenAI views it as a critical step toward developing more powerful AI systems.

"We’re sharing GPT-4.5 as a research preview to better understand its strengths and limitations," OpenAI wrote in a blog post. "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."