- Published on
2026 in LLMs (so far)
- Authors
- Name
- DeepHorus
“2026 in LLMs (so far)” is a year-in-review essay and annotated slide deck published on September 27, 2026, following the author’s closing keynote at the WeAreDevelopers World Congress North America in San Jose. The talk connects the past year’s developments into a chronological tour of the field, while companion materials on YouTube provide additional context. Rather than treating 2026 as a finished chapter, the article emphasizes that several months remain. Its central argument is that the most consequential change so far is not one isolated model release, but the emergence of coding agents capable of dependable, day-to-day software work.
For the author, the story begins in November 2025, when Anthropic released Claude Opus 4.5 and OpenAI introduced GPT-5.1. Both were broadly incremental improvements over earlier systems, but together with their agent frameworks they crossed an important practical threshold. Claude Code, which had existed since February 2025, and the younger Codex agent became substantially more reliable when paired with the updated models. Previously, agents often made errors that prevented routine use; by late 2025, the author judged the best combinations reliable enough to integrate into everyday development workflows.
That improvement prompted a deliberately ambitious experiment in 2026. After repeatedly failing to keep his New Year’s resolution of taking on fewer projects, the author decided to do the opposite and start as many as possible. His aim was not merely to automate familiar tasks, but to discover where the new agent technology stopped working by pushing it well beyond its apparent limits. The resulting workload included many projects in progress simultaneously, reflecting a broader shift among developers from evaluating AI tools in isolation to restructuring their personal processes around them.
The article also revisits the author’s deliberately absurd test for frontier models: generating an SVG of a pelican riding a bicycle. The benchmark cannot provide a comprehensive measure of intelligence, but it still combines several difficult requirements, including rendering two recognizable objects and depicting an improbable interaction. Results from November 2025 showed that progress had not eliminated every weakness. Claude still struggled to draw a convincing bicycle, while GPT-5.1 produced a poor bicycle frame despite its coding capabilities. The anecdote underscores that coding reliability and general visual reasoning do not necessarily advance at the same pace.
Another signal of the changing development landscape was the first commit to an obscure GitHub repository called “Warelay” in November 2025. Although the supplied notes do not yet explain the project’s eventual importance, the decision to highlight its obscure origin gives the year’s chronology a human dimension. The repository emerged during the holiday period, when individual developers were experimenting with the new Claude and GPT agent combinations. By January, those experiments had convinced many practitioners that the tools could support much larger ambitions than they had previously considered practical.
The author’s reactions were also shared during appearances and conference discussions. On the Oxide and friends podcast with Bryan Cantrill and Adam Leventhal, he predicted that it would become undeniable that LLMs write good code, that sandboxing would be solved, and that a major coding-agent security incident would eventually occur. The first prediction now appears substantially fulfilled, while sandboxing remains an active area of experimentation: about 40 of 277 sessions at the San Jose conference touched agent security or sandboxing. Rather than the catastrophic hijacking he anticipated, however, the year has produced considerable debate and noise without a clearly documented agent-driven economic disaster. His lighter predictions included comments from the Pope and an exceptional breeding season for New Zealand’s flightless kākāpō.
Looking ahead, the article presents the remaining months of 2026 as an open-ended test of agent limits. Coding agents are increasingly reliable, but security, containment, permissions, and failure recovery have not been fully resolved. The conference’s attention to sandboxing suggests the industry recognizes that greater autonomy must be matched by stronger infrastructure. Future progress will depend on whether developers can move from impressive demonstrations to controlled, repeatable systems, while continuing to expose the kinds of weaknesses revealed by the bicycle benchmark. The broader outlook is therefore cautiously optimistic: coding agents have crossed a meaningful usability threshold, but their long-term impact will be determined by the engineering and security practices built around them.
