OpenAI releases GPT-4.5 research preview, its largest model

OpenAI has released a research preview of GPT-4.5, which the company describes as "our largest and most knowledgeable model yet."

The announcement came in the GPT-4.5 System Card, the safety and capability documentation OpenAI publishes alongside major model releases. The system card format has become OpenAI's standard practice for disclosing evaluated risks and behaviors before or at deployment.

The release marks a notable step in OpenAI's model lineup. GPT-4.5 sits above GPT-4o in scale, according to the company's own characterization, and the "research preview" label signals that OpenAI is treating this as an early, limited rollout rather than a finished product. The company has used similar preview designations in the past when it wants real-world usage data before a broader launch.

Scale remains one of the central variables in frontier AI development. OpenAI's claim that GPT-4.5 is its largest model carries weight because raw parameter and compute scale — the approach behind GPT-3 and GPT-4 — is once again competing with efficiency-focused techniques such as distillation and mixture-of-experts architectures. A larger, more compute-intensive model typically means higher inference costs, which shapes who can access it and at what price.

The "most knowledgeable" framing points to breadth of world knowledge as the differentiator, distinct from reasoning ability that OpenAI has emphasized in its o-series models. For enterprise users and developers choosing between models, that distinction matters: knowledge-heavy tasks such as factual recall and domain expertise behave differently from step-by-step reasoning workloads.

The stakes extend beyond product strategy. System cards have become a focal point in the policy debate over AI transparency, with regulators and researchers scrutinizing how thoroughly companies document model risks before release. OpenAI's decision to publish the GPT-4.5 System Card alongside the preview keeps that disclosure practice in place for its largest model to date.

OpenAI has not stated in the released material how long the research preview will run or when a general-availability launch will follow. The company's track record suggests preview status serves as a data-gathering phase, meaning GPT-4.5's capabilities and safety profile may be refined before full deployment.