2025Model

GPT-4.5/o3: Next Generation Models

OpenAI releases GPT-4.5 and o3 series, continuously pushing LLM capability boundaries.

Overview

In 2025, OpenAI continued advancing LLM frontiers. GPT-4.5 further improved multimodal understanding and knowledge breadth. The o3 series inherited o1's reasoning capabilities, achieving breakthrough scores on the ARC-AGI benchmark, demonstrating progress toward AGI. These models reached new heights in code generation, scientific reasoning, and complex task handling. OpenAI's continuous innovation pushed the entire industry forward, while also triggering more discussions about AI safety and regulation.

Study Notes

Background, mechanics, impact, and key takeaways in one read

About 3 min read

Background & Context

In early 2025 OpenAI's product line reached a turning point. On February 27, GPT-4.5 (codename Orion) was released as a research preview: the last flagship of OpenAI's pure pretraining-scaling route, the largest model in parameters and training compute to date, achieving broader world knowledge, lower hallucination, and more natural conversation through scaled unsupervised learning, but without chain-of-thought reasoning, with Altman calling it "giant and expensive" and API pricing at 75 dollars per million input tokens. Meanwhile the reasoning line advanced: o3-mini shipped on January 31, and the full o3 on April 16, with breakthrough results on abstract-reasoning benchmarks like ARC-AGI. GPT-4.5 thus became the end of an era: the endpoint of pretraining scaling overlapping with the start of reasoning scaling.

Key Points

1GPT-4.5 (Orion) was released as a research preview on February 27, 2025, the last pure pretraining flagship.
2Scaled unsupervised learning gave broader knowledge and lower hallucination, but no chain-of-thought reasoning.
3API pricing was extreme (75 dollars per million input tokens), opening first to ChatGPT Pro users.
4o3-mini (January 31) and o3 (April 16) pushed the reasoning line to new heights.
5o3's breakthrough on ARC-AGI also exposed controversy over benchmark ceilings.

How It Works

GPT-4.5's technical significance lies in pushing pretraining scaling to the joint boundary of economics and technology: more compute, more data, and finer alignment, including scalable alignment techniques that train large models on data generated by smaller ones, bought intuition and knowledge, yet on math and code, which demand explicit reasoning, it still trailed o3-mini, directly proving that pretraining and reasoning are two complementary scaling axes. o3 represents the other axis: large-scale reinforcement learning that trains models to emit longer internal reasoning chains before answering and to allocate computation on demand at inference. On ARC-AGI, o3 scored about 75.7% at low compute and about 87.5% at high compute, far beyond prior models; but ARC-AGI is hard even for humans, and the benchmark quickly faced contamination and saturation controversy, again showing that reasoning benchmarks have shorter half-lives than model release cycles.

Impact & Legacy

GPT-4.5 and o3 together marked the 2025 industry consensus: capability scaling expanded from a single pretraining axis to the dual axes of pretraining plus inference-time compute, and every subsequent flagship reports both scaling curves. GPT-4.5's expense and short life, its API preview withdrawn in July 2025, showed that pure pretraining had approached its commercial limit of marginal returns, after which Altman announced GPT-5 would unify both routes. In the competitive landscape, o3's ARC-AGI results briefly ignited "AGI is near" discourse, then cooled under contamination and saturation critiques, while DeepSeek R1 simultaneously open-sourced much of the reasoning route's capability, pushing the cost curve down again. 2025 thus became the year of reasoning-model democratization, with thinking budgets, verifiable rewards (RLVR), and long chains of thought becoming standard configuration in every lab.

Key Takeaways

Pretraining scaling and inference-time compute are complementary axes; flagships must report both.
When pure pretraining's marginal returns approach zero, product form shifts quickly to reasoning models.
Abstract-reasoning benchmarks like ARC-AGI saturate and contaminate easily; breakthrough claims need methodological audit.
Reasoning-model competition quickly shifts from "can it think" to "how cheaply can it think", R1's shock.

Glossary

Test-Time Compute推理时计算

A new scaling axis investing more computation at inference, longer chains of thought and more samples, for stronger capability.

ARC-AGIARC-AGI

The Abstraction and Reasoning Corpus benchmark probing human-like few-shot induction, used as a touchstone by reasoning models like o3.

Research Preview研究预览

A release form that gives limited users access to a frontier model for feedback, as used for GPT-4.5.

Related Events

2024