OpenAI Makes GPT-6 the Default Model in ChatGPT

OpenAI started rolling out GPT-6 as ChatGPT's default model on October 7, 2026. Plus and Pro subscribers, along with Business and Enterprise plans, get GPT-6 Sol from October 7. Free and Go users get GPT-6 Luna starting October 8. OpenAI is launching GPT-6 together with a feature it calls Intelligent UI. A reply can now include a chart or buttons. It can also hold a form or a small tool that works inside the conversation. So answers in ChatGPT will start to look different. Codex users see no change. OpenAI says the models behind Work and Codex stay as they are. The Verge ↗

Anthropic Releases Claude Haiku 5.5, a Cheaper Small Model

Anthropic released Claude Haiku 5.5 on October 7, 2026, through its API and inside Claude Code. It is also on AWS, Google Cloud and Azure. Anthropic says it costs about 75% less to run than Haiku 4.5. That lowers the cost of apps built on Claude: $0.10 per million tokens sent in, roughly 750,000 words. But that price holds only while a request stays under 100,000 tokens. Above that line it rises fivefold, which VentureBeat says matters for long agent sessions. On OSWorld, a test of operating a desktop computer, Anthropic reports 72.4%, up from 15.7% for Haiku 4.5. Anthropic ↗

Microsoft Adds Built-In Limits for AI Agents on Windows 11

Microsoft made Execution Containers generally available on Windows 11 on October 7, 2026. The change is simple. Windows itself now blocks an AI agent from files and network connections it was not given, where until now that kind of limit depended on settings inside each agent. Anyone who lets an assistant work on their own computer gets a second line of protection. Support is already there in some agents. Microsoft's developer blog lists OpenAI's Codex and GitHub Copilot, as well as OpenClaw. Claude Code is coming. Windows Developer Blog ↗

Google Opens SynthID Detector to the Public

Google DeepMind opened SynthID Detector to the public at synthid.com on October 7, 2026. Sign-in works with a Google or Apple account, and OpenAI accounts work too. You upload an image or video, or an audio clip, and the site checks it for SynthID, an invisible watermark. It finds marks added by Google's and OpenAI's tools, and by those from NVIDIA and Kakao. Google says Apple's are coming. It cannot catch media from a tool that adds no mark, so a clean result does not prove a person made it. Google ↗

Google Releases Nano Banana 2.1 Image Model

Google released Nano Banana 2.1, its model for making and editing images, on October 6, 2026. It is in the Gemini app and Google Search, and in AI Studio and Flow. In Google's own ratings, people preferred it over both Nano Banana 2 and Nano Banana Pro. No independent tests were out at launch. Google's API docs list about 3.4 cents per standard image for developers, roughly half the earlier price. Gemini API docs ↗

HyperFrames Turns HTML Into Video

HyperFrames, from HeyGen, renders video from HTML. Its own description is short: write HTML, render video, built for agents. In other words, it is meant for AI agents that make videos by writing HTML. The TypeScript project gained 3,908 stars this week on GitHub's weekly trending list. GitHub ↗

Cursor Posts Its Plugin Specification

The cursor/plugins repository holds Cursor's plugin specification, which sets out how plugins for Cursor are defined. It also contains Cursor's official plugins. Developers can read the spec and use the official plugins as working examples. The TypeScript project gained 1,109 stars this week on GitHub's weekly trending list. GitHub ↗