The last six months in LLMs in five minutes

2026-05-19 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Intermediate, medium

Summary

The last six months in Large Language Models (LLMs), from November 2025 to April 2026, marked a significant inflection point, particularly for coding applications. November 2025 saw the "best" model title shift five times among Anthropic, OpenAI, and Google, with models like Claude Sonnet 4.5, GPT-5.1, Gemini 3, GPT-5.1 Codex Max, and Claude Opus 4.5. Crucially, coding agents, enhanced by Reinforcement Learning from Verifiable Rewards, transitioned from "often-work" to "mostly-work," becoming reliable daily tools. February 2026 saw the "Warelay" project evolve into "OpenClaw," a personal AI assistant, popularizing the generic term "Claws" and driving Mac Mini sales. February also brought Gemini 3.1 Pro, demonstrating remarkable image generation capabilities. April 2026 introduced Google's Gemma 4 series and GLM-5.1, a 1.5TB open-weight model, alongside Qwen3.6-35B-A3B, a 20.9GB open-weight model that outperformed Claude Opus 4.7 on specific benchmarks. The period's key themes are the dramatic improvement of coding agents and the unexpected performance of local, laptop-run models.

Key takeaway

For AI Engineers evaluating LLM integration, the rapid evolution from November 2025 to April 2026 means continuous re-evaluation of model choices is critical. You should now consider advanced coding agents as reliable daily drivers, significantly boosting productivity. Furthermore, explore capable open-weight models like Gemma 4 or Qwen3.6-35B-A3B for local deployment, as their performance now wildly outperforms prior expectations, potentially reducing reliance on costly cloud APIs.

Key insights

LLM capabilities, particularly coding agents and local models, advanced rapidly from November 2025 to April 2026, marked by frequent "best" model shifts.

Principles

• LLM performance leadership is highly volatile, changing hands frequently.
• Specialized training like RL from Verifiable Rewards enhances agent quality.
• Unconventional benchmarks reveal true model generalization.

Method

Reinforcement Learning from Verifiable Rewards (RLVR) was applied to increase code quality in LLM agents.

In practice

• Integrate advanced coding agents into daily development workflows.
• Evaluate open-weight models like Gemma 4 or Qwen3.6-35B-A3B for local deployment.
• Consider dedicated hardware like Mac Minis for running personal AI assistants.

Topics

• Large Language Models
• Coding Agents
• Open-weight Models
• Model Benchmarking
• Personal AI Assistants
• Reinforcement Learning from Verifiable Rewards

Code references

• bellard/mquickjs
• simonw/micro-javascript

Best for: NLP Engineer, CTO, VP of Engineering/Data, AI Engineer, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• The Only LLM Comparison Guide You Need in 2026 — In 2026, LLM selection shifts from "best overall" to "best for specific tasks, cost, and deployment."
• Programming with LLM Agents in 2025 — LLM agents, like Open Hands, significantly amplify developer productivity by automating iterative coding and visualization tasks.
• not much happened today — The AI landscape is rapidly evolving towards specialized, agentic systems, with open-weight models gaining significant ground against proprietary solutions.
• Running local models is good now — Local LLMs have matured significantly, now enabling effective agentic coding and deep model introspection.
• 🤖AI Agents Weekly: LLMs in 2025, YOLO in the Sandbox, Plan Caching for Agents, DeepTutor — Coding agents and reasoning models significantly advanced LLM capabilities and market impact in 2025.
• LLMs: A Journey Through Time and Architecture — LLM evolution prioritizes data quality, multi-stage pre-training, and architectural efficiency over sheer model size.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.