PyCon 2026 LLM Half-Year Review: Coding Agents and Claw Ecological Explosion

Core conclusion

At PyCon US 2026, well-known Python/LLM blogger Simon Willison used a 5-minute lightning speech to sort out the most important changes in the LLM field in the past six months. There are three core conclusions: coding agents have crossed the threshold of "daily usability", open source models perform amazingly on laptops, and the ecosystem of Claw-like AI assistants has exploded.

Key Points

• Time span: November 2025 - May 2026
• Coding agents move from "occasionally available" to "daily productivity tools"
• The title of "Best Model" changed hands 5 times between the three major manufacturers
• The performance of open source models (Llama 4, Qwen3.6, GLM-5.1) has been greatly improved on notebook-level hardware
• The OpenClaw ecosystem went from scratch to global attention in just 3 months

Background: Coding Agent’s “Aha Moment

If there is one point in time to define the past half year, it would be November 2025. Simon Willison calls it a pivotal moment in LLM’s history.

At that time, OpenAI and Anthropic had been using RLVR (reinforcement learning based on verifiable rewards) training models to improve code quality, in conjunction with the Agent framework of Codex and Claude Code. In November, these efforts came to fruition - the coding agent went from "occasionally available" to "daily driven." Developers can actually use it to get real work done and no longer need to spend most of their time fixing AI-generated errors.

In the past six months, the "best model" (based on subjective experience) has changed hands 5 times between different vendors:

Interestingly, Simon Willison used the strange test of "drawing a pelican and riding a bicycle" to compare various models - because a pelican is difficult to draw, a bicycle is difficult to draw, a pelican riding a bicycle is even more difficult to draw, and no AI laboratory will specifically train this task.

Claw Ecology: From Warelay to OpenClaw

In November 2025, a person named Pete submitted the first commit of the project "Warelay" on GitHub. Through December and January, the project went through multiple name changes, finally taking the world by storm in February as OpenClaw.

OpenClaw is defined as a "personal AI assistant", and Simon Willison observed that the common name for such tools has become "Claw" (including NanoClaw, ZeroClaw and other variants). He even made a funny joke: Mac Minis are being sold out in Silicon Valley because people are buying them to run their own Claws—"Mac Minis are the perfect fishbowl for Claws."

Simon prefers to use a more vivid metaphor to describe the Claw: It is like the mechanical claw of Alfred Molina's Doctor Octopus in "Spider-Man 2" - it is driven by AI and safe, but if the inhibitor chip is damaged, it will become evil and bite back on its owner.

Open source model: laptops can also run good results

April saw a real breakthrough for the open source model. The Llama 4 series is the strongest open source weighting model Simon has seen from a US company. Chinese AI lab GLM has unveiled GLM-5.1, a giant open-source model with 1.5TB of parameters—brilliant in capabilities but extremely demanding in terms of hardware.

Even more exciting is the release of Qwen3.6-35B-A3B by Qwen, a 20.9GB open source weight model that can be run on a laptop. Simon's actual testing found that its performance even exceeded Claude Opus 4.7 in the "Drawing Pelican" test.

Simon summarizes two major themes from the past six months:

1. Coding agents become really useful - from "occasional success" to "everyday tool"
2. Available notebook models significantly exceed expectations - Although far from cutting-edge models, the progress is amazing

Implications for content automation and making money with AI

This lightning speech has direct reference value for readers of WayToClawEarn:

• Coding agents are productivity tools: Claude Code and OpenClaw are no longer toys and can be used in real content production pipelines
• Open source models significantly reduce costs: Qwen3.6 This type of model can be run locally and is suitable for automation scenarios that require privacy protection or high-frequency calls.
• Agent framework is maturing: With DeepSeek, n8n and LangGraph, an end-to-end content automation system can be built
• Fast model change: Changing the list five times in half a year means that the strategy of relying on a single model is risky, and the multi-model scheduling strategy is more secure.

Related extended information

• Simon Willison ：The last six months in LLMs in five minutes
• PyCon US 2026 （）

Internal link guidance

• Want to learn how OpenClaw can be used for content automation? Watch: OpenClaw AI Agent Playbook: 500K TikTok Views in 5 Days, $588 MRR
• Want to build your own AI automated workflow? Reference: n8n + OpenAI
• Use DeepSeek to replace high-priced models: DeepSeek V4 vs Claude Code: 90% Cheaper, Same Quality