GPT-4.5
A large language model from OpenAI, released in 2025 and retired in 2026.
GPT-4.5 is a large language model from OpenAI, released on 27 February 2025 under the codename Orion. It was trained mainly through unsupervised learning and later removed from the API and ChatGPT as newer GPT models arrived.
- Released 27 February 2025 under the codename Orion
- Trained mainly with unsupervised learning plus human feedback
- Beat GPT-4o on MMLU across all 15 tested tongues
- API pricing of $75 per million input tokens in February 2025
- Withdrawn from ChatGPT on 27 June 2026
What it is
GPT-4.5 is a large language model built by OpenAI and released on 27 February 2025. It belongs to the GPT series and was known internally as Orion. The model was made available to paying ChatGPT subscribers on web, mobile and desktop, and it could also be reached through the OpenAI API and Developer Playground. It was designed to spot patterns, link ideas and produce creative text without relying on step-by-step reasoning. The model was later withdrawn from most services as newer GPT versions took its place.
How it came about
OpenAI released GPT-4.5 on 27 February 2025 for ChatGPT Plus and Pro subscribers. API and Developer Playground access ran until 14 July 2025. In April 2025 the company said it would retire the model from the API in favour of GPT-4.1. When GPT-5 arrived on 7 August 2025, GPT-4.5 disappeared from the API and from ChatGPT for Plus and Teams users, though Pro subscribers could still find it under Legacy Models. A later announcement on 28 May 2026 set 27 June 2026 as the date for full removal from ChatGPT.
Inside the model
Training relied mostly on unsupervised learning, a method that lets a model pick up patterns and connections from data without being told the right answer each time. That base was then refined with supervised fine-tuning, where humans provide example answers, and reinforcement learning from human feedback, which rewards responses people prefer. Microsoft Azure supplied the computing power for the training runs. Sam Altman called it a giant, expensive model, and the API pricing reflected that: $75 per million input tokens and $150 per million output tokens as of February 2025, against $2.50 and $10 for GPT-4o.
What it can do
On the MMLU benchmark, which covers 15 different tongues, GPT-4.5 beat GPT-4o in every one of them. Reviewers noted it handled subtext well and replied in a more natural, human-sounding way than its predecessor. Aaron Levie of Box told Axios the model was about 20% stronger than GPT-4o on certain jobs, such as pulling information out of very large datasets. It did not use chain-of-thought reasoning, the technique of working through a problem step by step before answering.
How to use it
The model reached users through ChatGPT on web, mobile and desktop for those on the Plus and Pro plans. Developers could call it through the OpenAI API and the Developer Playground until 14 July 2025. After GPT-5 launched on 7 August 2025, only Pro subscribers kept access, via a Legacy Models tab. A notice on 28 May 2026 said ChatGPT support would end on 27 June 2026.
Reception and debate
Most reviewers treated GPT-4.5 as a modest step up from GPT-4o rather than a major change. Cade Metz of The New York Times wrote that it marked the close of an era, being the firm's final chatbot without chain-of-thought reasoning, and the paper judged it unlikely to match the excitement around GPT-4. TechRadar's Eric Schwartz liked the grasp of subtext and the warmer tone, but said the jump from the earlier model felt small.
Why it matters
GPT-4.5 sits between GPT-4o and the later GPT-4.1 and GPT-5 releases, and its retirement shows how quickly OpenAI moved its users onto newer systems. Its API pricing, far above GPT-4o, also highlighted the cost of scaling up a model trained in this way.
Questions people ask about GPT-4.5
What is GPT-4.5?
It is a large language model from OpenAI, released on 27 February 2025 and known during development as Orion. It belongs to the GPT series and was offered to paying ChatGPT users and through the OpenAI API before being retired.
How was it trained?
Most of the work used unsupervised learning, letting the model find patterns on its own. That was followed by supervised fine-tuning and reinforcement learning from human feedback. Microsoft Azure provided the computing resources for the training.
How much did it cost to use through the API?
As of February 2025, the API charged $75 for each million input tokens and $150 for each million output tokens. For comparison, GPT-4o cost $2.50 and $10 for the same amounts.
Why was it retired?
OpenAI moved users to newer models. GPT-4.1 replaced it for API work from 14 July 2025, GPT-5 took over most ChatGPT access on 7 August 2025, and full removal from ChatGPT was set for 27 June 2026.
How did it differ from GPT-4o?
Reviewers saw it as an incremental upgrade. It handled subtext and natural tone better, and it beat GPT-4o on MMLU in every tested tongue, but it did not use chain-of-thought reasoning.
Good reading and sources
- Official websiteOfficial site
- OpenAI announces GPT-4.5, warns it’s not a frontier AI modelThe Verge, Article
- OpenAI launching GPT-4.5, its next general-purpose large language modelCNBC, Article
- OpenAI unveils GPT-4.5 'Orion,' its largest AI model yetTechCrunch, Article
The facts on this page come from the Wikipedia article "GPT-4.5", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.