In a closing keynote at the WeAreDevelopers World Congress North America in San Jose, developer Simon Willison reviewed the major large-language-model developments of 2026, publishing annotated slides and notes on his blog and a video on YouTube. According to Willison, the year effectively began in November 2025 with the release of Claude Opus 4.5 and GPT-5.1. He describes these as incremental improvements that nonetheless pushed their paired coding agents—Claude Code and Codex—from 'often make mistakes' to 'reliable enough to use on a day-to-day basis.' Willison notes that developers experimenting over the December holidays realized how much more the tools could do, and that 'be more ambitious' became a theme for his year. He recounts building a JavaScript interpreter in Python and a WebAssembly runtime in Python during a period he calls 'AI mania.' Willison also references predictions made on the Oxide and Friends podcast, including that it would 'become undeniable that LLMs write good code,' that sandboxing efforts would intensify (he counted roughly 40 of 277 conference sessions touching agent security), and lighter predictions about the Pope and New Zealand's Kākāpō parrots. He also tracks an obscure GitHub repository that renamed itself several times to become 'OpenClaw.' The account is one person's perspective, presented as a talk.
- Claude Opus 4.5 and GPT-5.1 released November 2025, improving coding agents to daily-use reliability per Willison
- Roughly 40 of 277 conference sessions touched on sandboxing or agent security
What it means for you
A well-known developer is arguing that AI coding tools crossed a line in late 2025 from 'makes too many mistakes' to 'reliable enough for daily work.' That's a genuine shift if you write software or hire people who do — but it's one person's read on the year, not a benchmark or study. For non-developers, this is mostly a signal that the tools are maturing, not a to-do list.
Try this
If you or your team write code, spend an hour this week trying a current coding agent (Claude Code or Codex) on a small real task and judge for yourself whether it's reliable enough to keep.
Who should care
Software developers, technical founders, and small businesses that build or maintain their own software or web tools.
Skip this if
You don't write software and don't manage anyone who does — this is a developer's year-in-review, not something that changes your plans.
Sources: Simon Willison — read the original