Monthly insight

October 2026

Published

The month in brief

Security dominated October. Researchers found AI coding agents leaking internal screenshots, and a backdoored Claude extension reached Cursor users. Claude Code also led a Stack Overflow survey of tool usage, and Mistral and Google made major launches.

Key developments

Security. Glow found more than 13,000 internal images from over 300 companies in public GitHub repositories. Coding agents had created them to host code-review screenshots, often through the open-source gitshot tool, which defaults to public repos. GitHub's CLI now has an --attach flag that removes the need for the workaround. Separately, a malicious version 2.1.313 of Anthropic's Claude extension sat on Cursor's marketplace for nearly 12 hours, affecting at most 5,000 users. CrowdStrike also tied attacks on South Korean banks to a suspect using Claude Code and the ARTEX pentesting agent.

Market share. According to the Stack Overflow survey as reported by Tech Insider, 65.5% of developers use Claude Code, against 58.7% for GitHub Copilot and 29.5% for OpenAI Codex. Totals exceed 100% because most developers run several agents.

Launches and funding. Mistral released Mistral Large 4, nicknamed "Le Chonk", an open-weight trillion-parameter model aimed at coding and cybersecurity, after a $3.3 billion round at a $24 billion valuation. Google Cloud previewed persistent Gemini agents with their own Gmail, Calendar and Drive. Pi 1.0, an MIT-licensed terminal agent from Earendil Works, claims about 40% fewer tokens through its Codemode tool calling. Manus raised over $500 million after its Meta deal collapsed.

New tools and ranking updates

Twenty tools were added on October 28:

  • Mistral: Mistral Large, Mistral Large 3, Mistral Medium, Mistral Small 4
  • OpenAI: GPT, GPT mini, GPT-5 Sol, GPT-5 Terra, GPT-5 Luna, gpt-oss-120b
  • Chinese labs: Qwen Plus, MiniMax M, MiniMax M2, GLM Flash, DeepSeek Pro, Kimi Pro, MiMo
  • Others: Muse Spark, Muse Spark Open, Claude Opus 4

The October ranking, published October 28, covers 148 tools and uses algorithm v7.10. That version reads technical and pricing data for tools with nested records, which several previously scored low on. On the site, the changelog also lists a weekly news digest and separate ranking boards for coding harnesses, models and productivity tools.

What to watch

Expect more scrutiny of what agents do unsupervised, from public repos to marketplace extensions. Also watch whether Google's Gemini agents leave private preview.