OpenAI releases a research preview of GPT-4.5, its largest and best model for chat, scaling up unsupervised learning rather than reasoning.
OpenAI released GPT-4.5 (codename Orion) as a research preview on 27 February 2025, first to ChatGPT Pro ($200 a month) users and developers, with Plus and Team the following week and Edu and Enterprise after that[1][2][3]
Sam Altman called it 'a giant, expensive model', said it would not be state of the art on benchmarks, and that OpenAI was out of GPUs and would add tens of thousands before the Plus rollout[2][3]
Reception called it an incremental upgrade and the last OpenAI chatbot without chain-of-thought reasoning[2]
Later: API access ended on 14 July 2025 in favour of GPT-4.1; it was removed from Plus and Team when GPT-5 launched (7 August 2025), and ChatGPT retirement was set for 27 June 2026[1][2]
Notable features
It is a general-purpose model trained mainly by scaling unsupervised learning plus supervised fine-tuning and RLHF; OpenAI said it feels more natural and presents inaccurate information less often than GPT-4o and o1 in one test[2][3]
It was trained on Microsoft Azure AI supercomputers by scaling up compute and data, and does not think before it responds, unlike reasoning models such as o1 and o3-mini[4]
OpenAI reports 62.5% accuracy on the SimpleQA factuality benchmark with a 37.1% hallucination rate, and human testers preferring it over GPT-4o on 56.8% to 63.2% of queries[4]
In ChatGPT it has search, file and image uploads and canvas but no Voice Mode, video or screensharing; in the API it supports function calling, Structured Outputs, streaming, system messages and image input[4]
API pricing at launch was $75 per million input and $150 per million output tokens, against $2.50 and $10 for GPT-4o; it was also offered in preview through Microsoft's Azure AI Foundry[2][3]
The first entry is always the pin's source. Overall confidence is a weighted average of how firmly each reference supports the start and end times used above; a reference counts half as much for every 180 days older than the newest.
web.archive.org· Published Feb 27, 2025· 5% of score
Archived copy of OpenAI's[1] GPT-4.5 announcement, which says it "was trained on Microsoft Azure AI supercomputers", gives SimpleQA accuracy of 62.5% and hallucination rate of 37.1%, and lists search, file and image uploads and canvas but not Voice Mode.
Suggest a correction
Something missing or wrong? Say it in your own words: a link that backs this pin up, a different start or end date and why, or a fact it lacks or gets wrong. The AI checks it against this pin's sources, searches for better ones, and adds any page that backs you up. The pin's own sources still count most.
OpenAI brings GPT-5.4 Thinking to ChatGPT, a frontier model combining GPT-5.3-Codex's coding with better work across tools, documents and spreadsheets.
