2026 in LLMs (So Far)

A rapid timeline traces how large language models crossed an invisible usability threshold in late 2025-early 2026: Claude Opus 4.5 and GPT-5.1 paired with coding-agent harnesses moved from error-prone helpers to reliable day-to-day tools. A running "Generate an SVG of a pelican riding a bicycle" benchmark illustrated incremental but meaningful improvements. Early November commits to a project called Warelay exploded into an OpenClaw phenomenon - 8,300 commits in under two months and now over 100,000 - sparking a new category of "Claw" personal agents (OpenClaw, NanoClaw, etc.), a surge in Mac Mini purchases to host them, and ephemeral social experiments like MoltBook that went viral then sank under spam before being acquired. Google’s Gemini 3.1 Pro also arrived with noticeably better image generation.

Those technical shifts produced cultural and operational consequences: an industry grappling with sandboxing and agent security, a sensation termed "Deep Blue" (AI-induced ennui among engineers), and an "AI mania" that encouraged ambitious, sometimes pointless projects (including micro-JavaScript and WebAssembly runtimes written in Python). Some organizations pushed agent-first engineering practices - StrongDM’s "Dark Factory" rules even mandate that code be written and not reviewed by humans - forcing new approaches to verification and trust. The year combined practical advances, infrastructure churn, intense experimentation, and rising questions about safety and governance.