OpenAI’s Secret Internal Model Rivals Top Programmers – GPT‑4.5 Unveiled

Beyond o1 and o3, OpenAI has an internal reasoning model that has not been publicly disclosed, revealed by CEO Sam Altman himself.

He said the model’s programming ability is currently in the global Top 50 and is expected to become number 1 by the end of the year.

Addressing the widely discussed compute issue, Altman stated:

Internally we have already reached GPT‑4.5, and moving to GPT‑5.5 will not require a hundred‑fold increase in compute.

The discussion, held at the University of Tokyo and hosted by President Teruo Fujii and Vice‑President Kaori Hayashi, also covered AI’s impact on education, the types of talent needed in the AI era, and various internal model updates.

Altman emphasized that humans are no longer superior to AI in tasks such as mathematics and programming, so cultivating creative vision and rapid adaptability is crucial.

Kevin Weil, OpenAI’s Chief Product Officer, encouraged everyone to start using AI in daily work and to solve new problems with it.

When asked about the potential of a hundred‑fold increase in compute, Altman explained that each GPT generation (1‑4) historically required roughly 100× more compute, but breakthroughs like the o1 series have dramatically improved efficiency.

He noted that the internal model is now at GPT‑4.5 level; under the previous scaling paradigm, GPT‑5.5 would need 100× more compute, yet new techniques reduce that requirement.

Altman listed recent progress in reasoning models: o1 ranks 9,800 in programming benchmarks, o3 has risen to rank 175, and the current internal model sits around Top 50, projected to reach the top by year‑end.

OpenAI aims to move toward open‑source development, focusing on small, powerful, and ultra‑fast inference models while continuing work on future GPT‑6/7 systems.

The company recently introduced a “Deep Research” feature for Pro users, an AI agent that can synthesize large amounts of online information and perform multi‑step research tasks.

Altman’s overarching goal for the year is to have a model that can answer extremely difficult questions after hours of reasoning.

When asked about open‑sourcing, Altman affirmed, “We will do it. We want to move toward open source,” though details on timing and scope remain undecided.

He also shared his views on frontier technologies such as brain‑computer interfaces, space engineering, and AI robotics, praising BCI companies as the most interesting he’s seen in six months.