2026 in LLMs (so far)

When coding agents struggle to complete tasks reliably, developers often abandon them for manual work. Throughout 2025, models like Claude and GPT showed promise but frequently stumbled on multi-step programming challenges. Then something shifted.

Training Approach

According to a keynote presentation at the WeAreDevelopers World Congress North America in San Jose, the year 2026 effectively began in November 2025 with two significant model releases: Claude Opus 4.5 and GPT-5.1. The presenter noted these releases followed a familiar pattern - incremental improvements over their predecessors rather than revolutionary leaps.

However, the presentation highlighted an important principle in language model development: models don’t improve uniformly across all capabilities. Instead, as models get better, they occasionally cross invisible thresholds where previously unreliable features suddenly become practical. The improvements in these November releases weren’t necessarily dramatic in benchmark scores, but they pushed certain capabilities over that critical line.

Notable Results

The specific capability that crossed the threshold was coding agents. Claude Code had existed since February 2025, while OpenAI’s Codex agent launched somewhat later. Both tools showed potential but lacked the reliability developers needed for production work.

When paired with the new Claude Opus 4.5 and GPT-5.1 models, these agents finally started working consistently enough for real-world use. The presentation framed this as a pivotal moment - not because the underlying architecture changed dramatically, but because the models reached sufficient capability to make agent-based coding practical rather than experimental.

This pattern of incremental improvement leading to sudden practical breakthroughs represents a recurring theme in LLM development. Small gains compound until previously marginal features become genuinely useful tools.

Running Locally

The presentation took a chronological approach, examining key trends throughout 2026 as they unfolded. While the talk covered the broader landscape of LLM developments through September 2026, the emphasis on coding agents in late 2025 and early 2026 suggests this capability shift drove significant adoption patterns.

The keynote was delivered as a closing address, attempting to tie together the year’s major developments into a coherent narrative. An annotated version of the presentation, including slides and detailed notes, was made available alongside a video recording at https://www.youtube.com/ for developers wanting to explore the timeline in depth.

Trade-offs

The framing of 2026 as beginning in November 2025 reflects how model releases don’t align neatly with calendar years. Major capabilities often emerge from models released months earlier, making it difficult to draw clean boundaries around “what happened when.”

The presentation’s focus on coding agents as the breakthrough application also highlights a trade-off in how we evaluate model improvements. Raw benchmark scores might show modest gains, but the practical impact depends entirely on whether those gains push specific use cases over the usability threshold. A model that’s 10% better overall but crosses that line for a critical application delivers far more value than the numbers suggest.

For developers tracking LLM progress, this creates a challenge: incremental improvements in model releases can mask significant shifts in practical capability. The gap between “this sometimes works” and “this reliably works” often appears smaller in benchmarks than it feels in production use.

The chronological exploration presented at WeAreDevelopers underscores how rapidly the LLM landscape evolved through 2026, with coding agents serving as an early indicator of models reaching new levels of practical utility.