2026 in LLMs (So Far) (simonwillison.net)

In 2026, significant advancements in large language models (LLMs) have been highlighted during the WeAreDevelopers World Congress, as the tech community analyzed key milestones and trends. Notably, the release of Claude Opus 4.5 and GPT-5.1 marked a pivotal moment for coding agents, transitioning from unreliable tools to reliable everyday assistants for developers. This improvement was underscored by a humorous benchmark involving LLM-generated graphics, where both models struggled with drawing complex items like a pelican riding a bicycle. The rise of personal AI agents, such as the OpenClaw, further demonstrated the demand for user-friendly AI that can autonomously perform tasks, marking a shift towards consumer-oriented LLM applications. The year has also seen transformational shifts in the use of AI in software development, exemplified by StrongDM’s radical approach of banning human-written code, relying instead on coding agents. AI has entered a phase of product market fit, with companies like Meta and Microsoft incorporating AI deeply into their workflow processes. Concurrently, there’s been a notable increase in the capabilities of open weight models, enabling sophisticated tasks even on personal devices, as evidenced by local models outperforming their more advanced counterparts. These developments not only define the current landscape of AI but also pave the way for the future role of LLMs in both personal and professional settings.