LLM half-year review: from the turning point of 2025.11 to the full availability of coding agents

Core conclusion

November 2025 has become a critical inflection point in the development of LLM. In the past six months, the coding agent has leapt from "occasionally available" to "daily workhorse", the performance of open source models has made a leap, and models on personal laptops have begun to surpass some cutting-edge models. For content creators and AI practitioners, this means that automated workflows have fully entered the practical stage.

Key Points

• Time of incident: November 2025 - May 2026 -Affected objects: content production teams, AI automation practitioners, independent developers
• Core changes: A leap in the practicality of coding agents + Explosive performance of open source large models + Personal devices can produce amazing results

Background and turning point: November 2025

November 2025 is a watershed in the development of LLM. Simon Willison systematically reviewed this process in a five-minute lightning talk at PyCon US 2026. In November alone, the title of "Best Model" changed hands five times between the three major vendors - from Claude Sonnet 4.5 to GPT-5.1, Gemini 3, GPT-5.1 Codex Max, and finally back to Anthropic's Claude Opus 4.5.

The bigger news is: Coding Agent just got better. OpenAI and Anthropic invested a lot of effort in RLVR (reinforcement learning based on verifiable rewards) throughout 2025, and the results of this work were concentrated in November. Coding Agents go from "occasionally usable" to "usable every day" - without spending a lot of time fixing their silly bugs.

Also in November, a then-little-known warehouse "Warelay" completed its first submission. The project, later renamed OpenClaw, took the entire developer community by storm in less than three months.

Key Impact

After February: OpenClaw Outbreak and Personal Device Counterattack

Entering February 2026, OpenClaw officially emerged as a "personal AI assistant" and spawned a number of imitators such as NanoClaw and ZeroClaw. The Mac Mini was once out of stock in Silicon Valley—someone joked that the Mac Mini is now the perfect fish tank for your Claw.

Also in February, Gemini 3.1 Pro was released, drawing the best content graphics yet. Google then released the Gemma 4 series—the best open source weighting models yet from an American company.

The surprise in April came from the Chinese team: GLM released the GLM-5.1 with 1.5TB parameters, while Qwen's Qwen3.6-35B-A3B only has 20.9GB, but can generate content beyond Claude Opus 4.7 on a personal laptop. **More than 90% API cost savings achieved through native or low-cost alternative models. **

Adaptation suggestions

What you can do immediately

• Embrace Agent-Driven Workflows: Coding Agents is good enough to become an everyday tool. If you are still manually debugging each AI output, you can now trust the Agent to complete the entire process from crawling to publishing.
• Build a multi-model strategy: Don't just rely on Claude or GPT. DeepSeek V4, Qwen, GLM and other Chinese models form strong competition in terms of cost performance and are suitable for use in different scenarios.
• Local models supplement the cloud: In content processing scenarios with high privacy requirements, locally available models such as Qwen3.6 are excellent supplementary solutions.

Recommended automation architecture

• Use n8n + Claude Code to build content production and quality verification pipeline
• Use OpenClaw to manage multi-Agent collaboration
• Use DeepSeek V4 to handle batches of high concurrent requests and reduce costs

#

# → Claude Opus 4.5

# → DeepSeek V4 / Qwen

# → Qwen3.6

# Agent → OpenClaw / Hermes Agent

Tool entry

Tools and frameworks mentioned in the text: OpenAI, ChatGPT, Claude, Claude Code, DeepSeek, Gemini, OpenClaw, n8n, Qwen, GLM, Hermes Agent.

Summary of six months

"Coding agents are becoming good enough to use, and models available for personal devices, while far weaker than leading-edge models, are starting to significantly exceed expectations." — Simon Willison, PyCon US 2026

For content automation and AI money-making direction, these changes mean: **AI automation in 2026 is no longer a question of whether to do it, but how to do it the fastest. **

Related extended information

• Simon Willison: The last six months in LLMs in five minutes
• Cursor Composer 2.5

Next action

• Want to use Claude Code to build your own automated content production line? Watch the tutorial: Claude Code automated writing practice: build an AI content production pipeline in 30 minutes
• Want to save 90% on API fees with DeepSeek? Watch the tutorial: Claude Code + DeepSeek V4 Building Tutorial: API Fees Dropped by 90% (15 Minutes)
• Someone has already successfully implemented it: Claude Code 48 hours to start a business: one person + US$29 monthly fee, monthly income in 3 months $9,000
• More exaggerated: Nat Eliason’s OpenClaw Experiment: $14,718 in 3 Weeks With a $1K AI Budget
• Newbies can read first: How to use OpenClaw + ChatGPT to build an AI automatic content rewriting system: from long articles to multi-platform adaptation in 30 minutes