Reasoning models were AI’s most important breakthrough of 2025, says DeepLearning.AI

In the year-end issue of its newsletter “The Batch”, published on 26 December, DeepLearning.AI named reasoning models such as OpenAI’s o1 and DeepSeek’s R1 the most transformative AI technology of 2025.

By “thinking step by step” before producing output, these models have significantly improved performance on complex maths, coding and scientific tasks. According to the report, reasoning models trained with reinforcement learning made notable gains: o1-preview outperformed its predecessor GPT-4o by 43 percentage points on competition maths and by 22 points on PhD-level science questions.

By the end of the year, Google’s Gemini 3 Pro, Anthropic’s Claude Opus 4.5 and GPT-5.2 were recognised as the top models for coding and agentic workflows.

The report also notes that China has accelerated domestic innovation in response to US chip restrictions. Huawei filed a record 37,000 patents in 2024 and invested USD 24 billion in R&D. In November 2025 Beijing also banned foreign AI chips in state-funded data centres — a move reflecting confidence in its domestic semiconductor industry.