GPT-4.5

What GPT-4.5 is

GPT-4.5 is a large language model released by OpenAI on 27 February 2025 and developed under the internal code name "Orion" ^[1]. At release it was offered to ChatGPT Plus and Pro subscribers, on the web, on mobile and in the desktop app ^[1]. The compute used to train it was supplied by Microsoft Azure ^[1]. Sam Altman described it as a "giant, expensive model" ^[1].

Price and scale

Pricing was set well above the model it followed. On the API, GPT-4.5 cost $75 per million input tokens and $150 per million output tokens, against $2.50 and $10 for GPT-4o ^[1]. Tokens are the small chunks of text a model reads and writes, and API prices are quoted per million of them, so these figures describe what a developer pays to send text to the model and get text back.

OpenAI presented the model as a research preview and described it as its largest and best model for chat at the time ^[4]. It also said that GPT-4.5 was more expensive than GPT-4o and was not a replacement for it ^[4].

How OpenAI described it

OpenAI attributed the model's behaviour to scaling unsupervised learning. In its account, learning from large amounts of data without task-specific labels let GPT-4.5 recognise patterns and draw connections without needing to reason step by step ^[4]. The company said early testing found that interaction with the model felt more natural, that its knowledge base was broader, that it followed user intent better, and that it showed greater "EQ"; it also said it expected the model to hallucinate less ^[4].

Reported benchmark results

OpenAI published comparisons with GPT-4o on four benchmarks. On GPQA, a set of graduate-level science questions, the new model scored 71.4% against 53.6% for GPT-4o. On AIME 2024, a competition mathematics test, it scored 36.7% against 9.3%. On MMMLU, a multilingual multiple-choice set, it scored 85.1% against 81.5%, and on MMMU, which poses questions about images, it scored 74.4% against 69.1% ^[4].

Reception

Cade Metz, writing in The New York Times, described the model as one that "signifies the end of an era" ^[1].

Lifecycle

GPT-4.5 remained a generally available product only briefly. OpenAI deprecated gpt-4.5-preview in the API on 14 April 2025, the same day GPT-4.1 launched, and shut it down on 14 July 2025 ^[2][3]. When GPT-5 launched on 7 August 2025 the model was removed from the Plus and Team tiers and kept only for Pro users as a legacy model ^[1]. Its removal from ChatGPT was announced on 28 May 2026, for 27 June 2026 ^[1].

Where it sits in the line

In OpenAI's sequence of models, GPT-4.5 followed GPT-4o and was itself followed by GPT-4.1 and GPT-5 ^[1]. Its release, its price and its short availability together mark it as a step that was measured against the cheaper model it came from rather than treated as the new default.