On Thursday, OpenAI officially released GPT-4.5, the model the company has reportedly been internally calling Orion, as a research preview to its Pro subscribers and developers using any of the API's paid tiers. OpenAI endorses GPT-4.5 as its largest model to date, as well as the best suited for chat. According to the company, GPT-4.5 boasts "deeper world knowledge" and "higher emotional intelligence" compared to previous models.

OpenAI also claims the new model hallucinates less and provides more natural, intuitive conversations with better understanding of human intent. This makes GPT-4.5 ideal for tasks related to writing enhancement, programming, and practical problem-solving. Notably, OpenAI performed some informal tests where it pitched its models against each other to showcase GPT-4.5's superior capabilities in creating SVG graphics and responding with greater emotional intelligence to personal situations.

Left: GPT-4.5, Center: GPT-4o, Right: o3-mini. Credit: OpenAI

Different models for different tasks

GPT-4.5's strengths and weaknesses compared to OpenAI's other models can be perceived by looking at the model's benchmark performance. While GPT-4.5 surpasses GPT-4o on several evaluations, it falls short of AI "reasoning" models from rivals like DeepSeek and Anthropic and of OpenAI's reasoning models (o1 and o3-mini).

On OpenAI's SimpleQA benchmark for factual accuracy, GPT-4.5 shows improvement over GPT-4o, o1, and o3-mini with a 62.5% accuracy score. Likewise, GPT-4.5 has a notably low 37.1% hallucination rate in the SimpleQA evaluation, followed by o1 with 44%, GPT-40 with 59.8%, and o3-mini with a staggering 80.3%. In contrast, while GPT-4.5 represents an improvement over GPT-40 on the SWE-Bench Verified evaluation (38% vs 3o.7%), it doesn't quite compare to o3-mini (high), as the latter scores 61% on the benchmark.

The Scaling Challenge

GPT-4.5 represents what OpenAI calls "a model at the frontier of what is possible in unsupervised learning." However, the release may also confirm the industry speculation that traditional pre-training scaling approaches are reaching their limits. OpenAI co-founder and former chief scientist Ilya Sutskever recently acknowledged this reality, stating, "We've achieved peak data" and "pre-training as we know it will unquestionably end."

OpenAI undoubtedly shares this sentiment, as CEO Sam Altman has confirmed GPT-4.5 will be the company's last (exclusively) non-reasoning model. The company plans to shift to offering integrated systems rather than standalone models. Altman has also confirmed that o3 will not be available as a standalone model but will be incorporated into GPT-5. The latter will ship as a system with several OpenAI technologies integrated into it, including o3.

Availability and pricing

Perhaps most notably, GPT-4.5 comes with steep costs. At $75 per million input tokens and $150 per million output tokens, it's dramatically more expensive than GPT-4o, which costs just $2.50 and $10 per million tokens, respectively. OpenAI acknowledges these high costs as a motivation for launching the model as a research preview. OpenAI claims it is "evaluating whether to continue serving [GPT-4.5] in the API long-term" and that feedback will be key in determining the future roadmap for this model.

ChatGPT Pro subscribers ($200/month) can access GPT-4.5 immediately, with ChatGPT Plus and Team users gaining access next week. Additionally, developers in all paid tiers can leverage GPT 4.5 in the Chat Completions, Batch, and Assistants APIs. According to Sam Altman, subscribers to the Pro and Plus plans were supposed to gain access to GPT-4.5 at the same time. However, due to OpenAI being out of GPUs, the company had to stagger the release as it scrambled to acquire even more GPUs.