← [[2026-10-09 Friday]] · [[Journal Index — 2026|Index]] ## Routine - [x] Open Daily Note - [x] **Daily Checks** - [ ] Self Care 🛀🧴 - [x] Pet Care 🐕🚶🏻‍♂️ - [x] Make Breakfast ☕🍳 - [x] Get Focused 🖥️ - [x] Check [Calendar](https://calendar.google.com) 📆 - [x] Check [Mail](https://mail.google.com) ✉️ - [ ] Review [[Yearly Log]] 📝 --- ## Log - --- ## Notes A month of AI model news, gathered this weekend. The story was price and agents that keep working on their own, not smarter models. - **Top-tier reasoning got cheaper for everyone.** OpenAI halved the price of its mid-size model and on 7 October put GPT-6 into every ChatGPT tier, free accounts included. Developers on Hacker News mostly called it a price cut rather than an improvement, and welcomed it anyway. OpenAI's own account is [GPT-6 and Intelligent UI for everyone](https://openai.com/index/gpt-6-for-everyone/). - **Agents now hold standing jobs.** OpenAI's Dots (29 September) run in the background with their own browser and report back when done. Anthropic's Opus 5.5 (22 September) is built to work unattended for hours, with a classifier checking each action before it runs. Anthropic describes that design in [Introducing Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5). - **A model that finds and patches security holes by itself.** Google's Gemini 4 Argon (30 September) is going to cyber defenders first, before developers or the public, which says something about what Google thinks it can do. Google's post is [Gemini 4 Argon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/). - **Frontier-class models you can download are coming from Europe and the United States.** Mistral, in France, promised the weights of its trillion-parameter Large 4 by the end of October; Reflection AI, in the US, promised Beam under an open licence this month, and Nvidia is reportedly in talks to buy the company. Their own write-ups: [Introducing Mistral Large 4](https://mistral.ai/news/mistral-large-4/) and [Introducing Beam](https://reflection.ai/blog/introducing-beam). - **Very cheap "good enough" models are changing how developers work:** a cheap model does the routine work and an expensive one reviews it. A developer's essay about DeepSeek's latest, [Why isn't the industry freaking out about DeepSeek 4.1 Flash?](https://www.dgt.is/blog/2026-10-07-deepseek-freek-out/), started that argument this week. - **Big models now run on a gaming PC.** An open-source project runs Alibaba's 125-billion-parameter Qwen model on a single consumer graphics card, with nothing leaving the machine. The project is [Strata, on GitHub](https://github.com/Niko1221/Strata). - **The plumbing matured.** The Model Context Protocol, the open standard that lets AI tools plug into other software, finished its 2026 revision this summer; the details are on [the protocol's own blog](https://blog.modelcontextprotocol.io/).