GPT-4.5 'Orion': OpenAI's last model without chain-of-thought
In one sentence OpenAI releases GPT-4.5 (internal name Orion) as a research preview: a large model trained mainly by scaling unsupervised learning, with no chain-of-thought reasoning. It hallucinates less than GPT-4o but costs $75/$150 per million tokens in the API.
On February 27, 2025 OpenAI releases GPT-4.5, internally called "Orion". It is a very large model, improved mainly the classic way: more data and more compute at training time, without the step-by-step "reasoning" phase of the o-series models.
OpenAI presents it as its best model for conversation: broader knowledge, more natural writing, fewer made-up answers. Sam Altman called it a "giant, expensive" model. The price confirms it: in the API it costs $75 per million input tokens and $150 per million output tokens, versus $2.50 and $10 for GPT-4o (30x and 15x as much).
A few weeks earlier Altman had announced that GPT-4.5 would be OpenAI's last model without chain-of-thought, and that the next step would be merging the GPT and o series into GPT-5.
Companies
OpenAI
Tools
GPT-4.5
Tags
Sources