OpenAI spent Monday explaining how it will mark its own text for European regulators and how it will sell ads next to pictures ChatGPT draws. API customers can opt in to a hidden watermark called textGrain. Eligible ChatGPT and Codex chats in the EU get the same treatment in the coming weeks. A Brooklyn lab of former DeepMind researchers posted a 501-billion-parameter open-weight model the same day, with weights still later this month.
In the news
OpenAI textGrain: EU watermark, API opt-in
The EU AI Act requires generated text to be machine-identifiable. OpenAI's answer, published Monday, is textGrain: an invisible statistical signal in word choice, plus a detector that reports whether an OpenAI watermark is present rather than who wrote the passage.1 API customers globally can opt in on select models, with watermarking off by default. In the coming weeks, eligible ChatGPT and Codex text in the EU, on all plans, will carry the signal. That rollout is not a global default.
OpenAI published its own limits. At a 1 percent false-positive rate, detection sits around 80 percent on 200-token psychology passages and around 95 percent at 400 tokens. Math is much worse: replacing 10 percent of the words with synonyms dropped 400-token detection from about 92 percent to 66 percent, and swapping a quarter of the words knocks it to 17 percent. Applications for the detector are open to approved researchers and expert organizations. Image and audio verify tools stay public, and OpenAI says it will open-source the technique later.1
ChatGPT visual ads during image generation
OpenAI is testing a labeled visual ad that appears while ChatGPT is generating an image, kept in a slot next to the picture rather than inside it.2 The test starts later this month in the United States with an initial advertiser group. The company repeats that ads do not influence answers. It claims 1.2 billion people use ChatGPT each week. That reach figure is OpenAI's.
Advertisers sign up at ads.openai.com. OpenAI is wiring conversion and attribution partners for the test, and it says brand-suitability pilots will not see private conversations. Partner conversion numbers from WeightWatchers, Dose, and Portland Leather are advertiser PR, not an independent audit.
Reflection Beam: 501B open-weight, later this month
Reflection, a Brooklyn lab of former DeepMind researchers, posted its first open-weight model. Beam is a sparse mixture of experts with 501 billion total parameters and 23 billion active, pretrained on 23.8 trillion tokens, text-only, aimed at coding, reasoning, and agents.3 Weights, the technical report, a model card, and developer artifacts come later this month. Early access is on platform.reflection.ai while the lab finishes red-teaming. You cannot pull the weights today.
The company table lists SWE-Bench Pro v2-Hard at 77.2, Terminal-Bench v2.1 at 80.1, AIME 2026 at 97.8, and GPQA Diamond at 90.5. Reflection claims parity with GLM-5.2 on advanced reasoning at three to four times less inference compute. TechCrunch reports those scores are not independently verified, and it cites a one-million-token context window.4 The lab describes a reinforcement-learning run of more than 100 million rollouts on 10,500 NVIDIA GB300 GPUs over four weeks.
