News Desk

OpenAI Adds Hidden Watermarks to ChatGPT Text in the EU

Grok (X search) · Perplexity API · GitHub Trending · Claude Opus 5.5 (writer pass)

Get News Desk by email

One email a week: the AI stories that mattered, explained simply.

In OpenAI's own tests, swapping a quarter of the words for synonyms cut detection of the mark from about 92% to 17%.

Also today: Reflection names the free AI model Axios reported on yesterday, Devin starts keeping notes between sessions, and a free video model that adds sound arrives.

Top story · Policy

OpenAI said on October 5, 2026 that text from ChatGPT and Codex, its coding assistant, will carry an invisible watermark called textGrain for EU users, rolling out over the coming weeks. The mark is a statistical pattern in word choice. It answers the EU AI Act, which requires AI-generated text to be machine-readable as such. If you use ChatGPT in the EU, nothing changes on your screen. OpenAI says the mark does not identify you. For now, only approved researchers have the detector. Outside the EU, ChatGPT is unchanged, and developers anywhere can switch the mark on through OpenAI's API, where it is off by default. OpenAI's announcement ↗

Hardware

Update: Reflection Announces Beam, a Free-to-Download AI Model

Reflection AI, the Nvidia-backed US startup Axios reported on October 5, announced Beam that evening. It is a coding and agent model whose files Reflection says anyone may download and reuse under the Apache 2.0 license later in October. Beam is still in safety testing; early access is by waitlist. Reflection's scores, unchecked outside the company, place it near GLM-5.2, a Chinese open model from June, and behind newer ones. Few people can run it at home. Beam has 501 billion parameters, the internal values that hold what a model learned. Reflection's announcement ↗

Agents

Cognition's Devin Now Remembers Lessons Between Sessions

Cognition added two features to Devin, its coding agent, on October 5, 2026. Memory keeps your preferences and project lessons from one session to the next. Dreaming is a nightly cleanup that merges duplicate notes and deletes stale ones. The notes are plain text files stored in Git, and Cognition published the format as an open spec, Agent Memory Repo, so tools such as Claude Code or Cursor can use the same layout. For a user, that means fewer repeats of the same house rules every morning. Cognition's write-up ↗

Video

Kandinsky Lab Releases Free Video Model With Sound

Kandinsky Lab published Kandinsky 6.0 Video on October 6, 2026, with files anyone can download and use under the MIT license. It turns a text prompt or a still image into a five-second clip. Each clip comes with audio, including lip-synced speech. There are two sizes, Lite and Pro. You do not need a powerful computer to try it: a free demo is up on Hugging Face, and the hosting service fal runs both versions. In Kandinsky Lab's own human ratings, people preferred Pro's speech over LTX 2.5, another open video model. Kandinsky 6 on GitHub ↗

Pricing

Tester Measures What ChatGPT Pro Buys in Codex

InventiveHQ published measurements on October 4, 2026 of how much Codex use ChatGPT's Pro plans include. OpenAI is halving the $200 plan's Codex allowance on October 30. Under the new limit, InventiveHQ measured a full week of use at about $640-670 at OpenAI's pay-as-you-go developer prices, roughly 14 times what the plan costs per week. The $100 plan came to about $330-380. OpenAI has not published comparable figures. Read more ↗

Benchmarks

Tokenless Reports 110 Grading Bugs in AI Coding Tests

Tokenless, a startup, said on October 5, 2026 that its EnvCheck tool found 110 problems in how three common coding tests grade AI models. DeepSWE had the most, with 70. Terminal-Bench had 24, and BFCL had 16. The company has filed them with the test makers. These tests produce the scores labs quote when they launch models, including Beam's numbers above. Read more ↗

Trending on GitHub

Open source

ponytail Tells Coding Assistants to Write Less Code

ponytail gained 8,596 stars on GitHub this week, the most on the weekly Trending list. It is a set of instructions for AI coding assistants. They push the assistant to reuse what a project already has and to add as little new code as it can, which leaves less for a person to review. Read more ↗

Tools

tuios Runs Several Coding Assistants in One Terminal

tuios gained 703 stars this week. The free tool splits a single terminal window into panes so several coding assistants can work side by side, and sessions survive a restart. A shared inbox shows which assistant is waiting on you. Read more ↗

Infrastructure

Cloudflare Publishes cloudflare-os, a Workspace for AI Agents

Cloudflare's cloudflare-os gained 663 stars this week. It is a workspace that runs on Cloudflare's servers, where AI agents write documents and build small apps. Agents can also work with a company's own data there. Teams set it up on their own Cloudflare account. Read more ↗

Google moves free Gemini users to its smaller Flash-Lite model on October 9.

Sign in

Sources

Get News Desk by email

One email a week: the AI stories that mattered, explained simply.

Build something with AI

Get the 16 prompts we use to plan, build and launch real projects with AI. Free.

Get the free prompts →