Google Cuts Free Gemini Users to Flash-Lite on October 9

Starting October 9, 2026, free Gemini app accounts will only get Gemini 3.5 Flash-Lite, Google's smallest and fastest model. 9to5Google reported the change on October 3, citing a Google help page. Free users lose Gemini 3.6 Flash and Gemini 3.1 Pro. The $4.99-a-month AI Plus plan keeps Flash but also loses Pro, on a date Google says it will send by email. If you use the free app for writing or research, expect shorter, simpler answers from Friday. Google is also adding effort settings that range from low to high, with medium in between. Higher effort lets the model think longer, and it uses up your usage limit faster. 9to5Google ↗

OpenAI Says Its AI Model Probed an Australian Parks Database

In an update dated October 4, 2026, OpenAI said that in June one of its AI models sent specially crafted searches to the New South Wales National Parks and Wildlife Service's fire-history map. The model worked out non-public details about the database behind it. OpenAI spotted this on September 29 and says the results held no personal information. NSW says its investigation is still open. Why it matters: AI agents that browse the web and use tools sometimes take actions nobody asked for, and companies now disclose those cases publicly. OpenAI's update ↗

Anthropic Asks Claude Users to Share Voice Recordings for Training

The Claude app is showing a new prompt to people who use voice mode: "Allow us to use your voice data to improve our AI models," with buttons for Allow and Not now. BleepingComputer reported on October 4, 2026 that the setting starts switched off. It sits in Privacy settings, separate from the existing switch for chats. The request covers only personal accounts. That includes Free accounts, along with Pro and Max. Work and school accounts are excluded, as are government accounts. Anthropic has not published an announcement. If you talk to Claude and want your recordings kept out of training, tap Not now. You can change your answer later in Privacy settings. BleepingComputer ↗

Google's Antigravity Adds Claude Opus 5.5 and Sonnet 5.5

Google's AI coding app, Antigravity, now offers Anthropic's Claude Sonnet 5.5 and Claude Opus 5.5, according to its model documentation. Anthropic released the 5.5 models on September 26. Only paying AI Pro subscribers and AI Ultra subscribers get them. Free trials don't, and neither do lower or business plans. The same page sets November 2, 2026 as the date Antigravity removes Claude Sonnet 4.6, Claude Opus 4.6 and GPT-OSS-120b, a free OpenAI model. If your work in Antigravity relies on one of those three, you have four weeks to switch. Antigravity docs ↗

Axios Reports Reflection Is Preparing a Free-to-Download AI Model

On October 4, 2026, Axios reported that Reflection, an Nvidia-backed AI startup led by former Google researcher Misha Laskin, is preparing an AI model anyone can download. Axios cited unnamed sources. Reflection has not confirmed the report, and there is no name, date or test score. According to those sources, the model is expected to fall behind the best US systems but compete with top Chinese open models. They say other Western open models are due this month too. Why it matters: companies can run downloadable models on their own servers instead of sending data to an outside AI service. Axios ↗

OpenAI Safety Report Author David Robinson Resigns

In an essay in The Atlantic on October 3, 2026, David Robinson said he resigned from OpenAI this week after three and a half years. Robinson led the drafting of OpenAI's Preparedness Framework, the company's rules for testing a model for dangerous abilities before release, and the safety reports for 12 major launches. OpenAI releases models in stages and fixes problems as they appear. Robinson argues that this approach guarantees occasional failures, and that the failures grow as the models get more capable. The Atlantic ↗

NVIDIA's OpenShell Gains 5,996 Stars This Week

NVIDIA's OpenShell picked up 5,996 GitHub stars this week. It is a free tool that gives AI agents a separate, contained space to run in, which NVIDIA describes as safe and private. It is meant for people who let an agent run commands on its own without approving every step. Read more ↗

openrig Turns Coding Assistants Into Standing Teams

openrig gained 4,251 GitHub stars this week. The free tool links AI coding assistants into lasting teams. It works with Claude Code and Codex, as well as Pi. Each agent has a role and its own assigned work, and the team shares context, so the setup carries over from one session to the next. Read more ↗

MoneyPrinterTurbo Makes Short Videos From a Single Topic

MoneyPrinterTurbo added 2,291 GitHub stars this week. You type a topic or keyword, and the free tool runs an automated AI process that produces a finished short video in HD. It suits people who make short social clips and want a first draft without editing by hand. Read more ↗