2025: The Year in LLMs (simonwillison.net)

In 2025, the landscape of Large Language Models (LLMs) evolved significantly, marked by advancements in reasoning capabilities and the advent of coding agents. Following OpenAI's introduction of Reinforcement Learning from Verifiable Rewards (RLVR) in late 2024, many leading AI laboratories launched their own reasoning models, which excel at complex tasks, including problem-solving and debugging code. These models can intelligently plan and execute multi-step tasks by utilizing tools, demonstrating their value in applications such as enhanced AI-assisted search and sophisticated code generation. For example, coding agents like Anthropic’s Claude Code and OpenAI's Codex allow for asynchronous task handling, where users can initiate coding projects and receive completed solutions without manual oversight. The significance of these developments lies not just in technical capabilities, but also in their implications for the industry. With subscription models now reaching as high as $200/month, there is a demographic shift towards users willing to invest substantially in advanced AI tools. This year also highlighted the potential risks associated with deploying LLMs, such as the normalization of risky behaviors in software usage. Overall, 2025 has ushered in a new era for LLMs, characterized by increased functionality, broader applications, and emerging trends that will define the future trajectories of AI and machine learning.