On September 27, 2026, Simon Willison published 2026 in LLMs (so far) — an annotated slide deck and essay originating from his closing keynote at the WeAreDevelopers World Congress (WWC26) North America in San Jose.

Known across the developer ecosystem for his hands-on, pragmatic evaluations of generative AI, Willison compiles nine months of breakneck progress into a cohesive timeline of model breakthroughs, security inflection points, and shifts in day-to-day software development.

Key Highlights from Willison’s Rundown

1. The Threshold of “Usable Every Day”

Willison traces the real starting line of 2026 back to late 2025 releases—specifically Claude Opus 4.5 and GPT-5.1 paired with harnesses like Claude Code and OpenAI Codex. Rather than just raw benchmark increases, these models crossed an invisible utility line: coding agents shifted from error-prone novelties into tools reliable enough for production work.

2. Willison’s Famous “Pelican Riding a Bicycle” Benchmark

Throughout his testing, Willison continued his visual sanity check: asking models to generate raw SVG code for a pelican riding a bicycle. The test tracks:

  • Spatial reasoning and anatomy understanding
  • Multi-object composition (pelican + mechanical frame)
  • Vector syntax formatting without visual feedback

While early 2026 models produced warped frames and disembodied beaks, recent iterations illustrate noticeable leaps in spatial coordination.

3. Agent Sandboxing and the Security Frontier

A central theme of the presentation centers on agent safety and runtime sandboxing:

  • Ecosystem Focus: Around 40 out of 277 conference sessions at WWC26 touched on sandboxing or agent authorisation.
  • Rogue Agents & Isolation: Isolating LLM execution environments (containers, WebAssembly/MicroPython environments, restricted API tokens) has become the defining operational problem as models gain file system and network execution capabilities.

Primary Links & Resources

  • Original Annotated Essay & Slides: Simon Willison’s Weblog (Sep 27, 2026)
  • Conference: WeAreDevelopers World Congress (WWC26) North America, San Jose
  • Related Tools Mentioned: Claude Code, Codex Remote, Datasette