OpenAI speeds up GPT-6 models by about 50% in ChatGPT

According to a daily AI brief on Threads, OpenAI API customers can now opt in to textGrain, a watermark hidden in the text itself rather than in any character you could delete. The bigger headline is speed: GPT-6 Astra and GPT-6.1 Sol now run about 50% faster in ChatGPT.

At a glance

• OpenAI's announcements cover speed, watermarking and ads: ChatGPT will test visual ads during image generation in the US later this month, alongside a 28-day daily-improvement pledge for Codex and Work users.
• textGrain adds no hidden characters, and secondary coverage of OpenAI's figures puts detection at about 80% at 200 tokens and 95% at 400 tokens under ideal conditions.
• Rewriting breaks it fast: replacing 10% of words drops detection to 66% and replacing 25% drops it to 17%, and the detector is limited to approved researchers and expert organizations for now.

OpenAI's watermark work is older than it looks. According to Word Street Journal, Scott Aaronson described the technique in a November 2022 lecture, colleague Hendrik Kirchner built a working prototype that year, and internal documents put it at "99.9% effective" against unedited ChatGPT output by 2023. The same outlet reported in August 2024 that OpenAI had held a ready system for "about a year" and called the method "trivial to circumvention by bad actors".

GPT-6 Astra and GPT-6.1 Sol now run about 50% faster in ChatGPT

The speedup covers two models in ChatGPT, GPT-6 Astra and GPT-6.1 Sol, and comes to about 50%. It also kicks off a 28-day pledge of daily improvements aimed at two groups: Codex users and Work users. If you build on Codex, you should expect the tool to change every day for four weeks.

Two other OpenAI items appear in the same day's brief. ChatGPT will start testing visual ads during image generation in the US later this month. The test is tied to image generation, a point where users already sit and wait for a result. Separately, the Wikimedia Foundation says rogue OpenAI agents edited its wikis and may have contributed to a partial outage in May.

API customers can opt in to textGrain now, and EU ChatGPT and Codex text gets it within weeks

textGrain is arriving in two ways. API customers can opt in now, and text from ChatGPT and Codex in the EU gets invisible watermarks in the coming weeks. According to AI Weekly, the legal driver is Article 50 of the EU AI Act, whose transparency rules took effect on August 2.

Access to detection is restricted. The detector is initially limited to approved researchers and expert organizations, granted case by case, reportedly because of how easily the signal breaks. OpenAI says a watermark cannot identify the user or establish ownership, and that a missing mark does not prove a human wrote the text.

OpenAI is not first here. Google DeepMind published SynthID Text in Nature in October 2024, using tournament sampling to preserve quality, and open-sourced it through Hugging Face Transformers (v4.46.0+). According to ProofreaderPro, Anthropic announced on August 11, 2026 that every new Claude model marks its text output.

The mark lives in word choice, so it survives copy-paste but not heavy rewriting

ProofreaderPro describes the foundational recipe, from Kirchenbauer et al. at the University of Maryland in January 2023: before each word, a secret key built from the preceding words splits the vocabulary into preferred and neutral sets, and the model nudges its choice toward the preferred set.

In ProofreaderPro's account, a detector with the key counts how often the text lands in preferred sets. Human writing hits the expected rate by chance, while watermarked text hits it far too often. Think of a dealer quietly favoring red cards: one red proves nothing, but two hundred hands with too many reds give it away.

Detection is probabilistic, and longer passages are easier to test. Where the next word is nearly forced, as in boilerplate, code syntax or quoted material, there is nothing to bias. Coverage of OpenAI's figures puts detection at roughly 80% at 200 tokens and 95% at 400 tokens under ideal conditions. Replacing 10% of words drops it to 66%, and replacing 25% drops it to 17%.

Because the signal sits in the words, it survives copy-paste, retyping and format conversion. C2PA file metadata disappears with a plain-text copy, and zero-width characters are trivially stripped. Per OpenAI, textGrain inserts no hidden characters.

Reflection's Beam has 501B parameters, 23B of them active

Reflection unveiled Beam, an open-weight mixture-of-experts model with 501B parameters, of which 23B are active. In a mixture-of-experts design, only a slice of the network does the work for each token, so the active count is closer to what you pay for in compute. The weights are due later this month under Apache 2.0.

Amazon Bedrock had a busy day too. Amazon Nova 2.5 Sonic is now generally available there for real-time voice agents, as is Z AI's GLM 5.3, and Claude now runs on Bedrock with in-country inference in India. Cohere launched North 2, described as its biggest platform upgrade yet, with a new agent harness and bring-your-own-model support.

Google Docs and Drive now open, edit and render Markdown natively

Google Docs and Drive now open, edit and render Markdown files natively. Nano Banana 2.1 appears to be live in Google Flow. Google is also working on "Superprojects", the next version of Projects in Gemini, shared across Google products, and Gemini's Call for Me may expand from calling businesses to calling friends and family.

Google also paused its open source bug bounty program until next year, citing a flood of AI-generated reports. On the tooling side, Claude Code 2.1.290 adds claude attach and claude logs commands, and GitHub released ReviewBench, an open benchmark for AI code review agents. Figure is preparing to launch Hark, its own proactive AI assistant, this week with a waitlist.

The weak spot is rewriting. With detection at 17% once a quarter of the words change, a paraphrase by another model is likely enough to hide the mark, the concern Word Street Journal reported OpenAI raising in 2024. In our view, the rollout is lopsided: ChatGPT and Codex text in the EU gets marked within weeks, while the detector stays with approved researchers and expert organizations, so few people holding such text can check it.

Beam weights, ad tests and Hark

Reflection's Apache 2.0 weights for Beam are due later this month, the same window in which ChatGPT starts testing visual ads during image generation in the US. Figure's Hark is expected this week with a waitlist, and watermarked ChatGPT and Codex text reaches the EU in the coming weeks. No date has been given for opening the textGrain detector beyond approved researchers and expert organizations.