Altman's roadmap post calls GPT-4.5 OpenAI's last non-chain-of-thought model
On 2025-02-12 Sam Altman posted on X that GPT-4.5, internally Orion, would be OpenAI's "last non-chain-of-thought model," and that GPT-5 would be a system integrating much of OpenAI's technology,…
- Date
- 12 February 2025
- Who
- OpenAI
- People
- Sam Altman
- Confidence
- High on the post's wording (as quoted in Willison's archive); Medium on GPT-4.5
- Deep dive
- Reasoning II, from o1 and o3 to DeepSeek-R1 and the labs that replicated them
Tier: Supporting · Significance: 3/5 · Org(s): OpenAI · People: Sam Altman · Confidence: High on the post's wording (as quoted in Willison's archive); Medium on GPT-4.5 details (Wikipedia pointer) On 2025-02-12 Sam Altman posted on X that GPT-4.5, internally Orion, would be OpenAI's "last non-chain-of-thought model," and that GPT-5 would be a system integrating much of OpenAI's technology, including o3, which would no longer ship as a standalone model (post as quoted in Willison's archive). GPT-4.5 launched on 2025-02-27 for Plus and Pro subscribers at $75 / $150 per million input / output tokens against GPT-4o's $2.50 / $10, was reviewed as an incremental step, and lost API access on 2025-07-14 (Wikipedia, a pointer). The post matters here because it is the primary evidence that the o-series and the GPT-series were meant to merge, which is the origin of the GPT-5 router that arrived 176 days later, and GPT-4.5 is the best-documented case of a large pretrained model priced against RL-trained reasoning. I infer that its price and muted reception strengthened the case for spending on RL and inference compute, though OpenAI has not said so. Sources: Willison, 2025-02-12 · Wikipedia, GPT-4.5