GPT-4.5 'Orion': OpenAI's last pure pre-training model
In one sentence OpenAI releases GPT-4.5 (codename Orion) as a 'research preview'. The largest model the company ever trained with traditional scaling, but expensive — marking the end of the pure pre-training era.
OpenAI releases GPT-4.5, codename "Orion". It's a very large model trained the traditional way (pre-training on huge text corpora). Sam Altman describes it as a model "that feels like talking to a thoughtful person" — fewer errors, fewer hallucinations, better writing.
But: it's extremely expensive to use (75x pricier than GPT-4o), and benchmark improvements are modest. The interesting part is what they don't say: Orion is probably OpenAI's last model built purely on "more data + more GPUs". From here on, the industry shifts to reasoning (o1, o3) and inference-time scaling, not just training scale.
Companies
OpenAI
Tools
GPT-4.5
Tags
Sources