It’s 2026, and looking back at how LLMs developed in 2025, I can clearly feel that AI finally stopped just talking and started doing things.
Simon Willison’s year-end review is excellent. Condensed into one line, his take on how LLMs changed in 2025 is roughly: Reasoning + agents + long tasks moved LLMs from tools to colleagues
A few points that resonated with me too:
- Reasoning is no longer just a term on benchmarks; it directly affects whether complex tasks get done
- Coding agents and CLI agents have genuinely entered engineers’ daily work, not just demos
- Starting with Codex in the middle of the year, LLMs can work continuously for hours, breaking down problems, fixing their own mistakes and continuing unfinished tasks
- $200/month subscriptions started to make sense (meaning people are really using them to make money or save time), though I’m more of a $20 × n person
Looking back, models advanced by leaps and bounds in 2025, but what’s really worth noticing is that AI finally started fitting into how people work.
If you also had a feeling this year of “how did so many things suddenly become things I can hand to AI,” you’re probably already standing at this turning point. It also means that future competitiveness won’t just be about whose model is stronger, but about who can put LLMs into the right work systems to create long-term value.
The original is long, and I strongly recommend reading it.