OpenAI: everyone gets GPT-6 in ChatGPT, with interactive answers. On Wednesday, OpenAI began rolling out GPT-6 in the ChatGPT chat interface: Plus, Pro, Business and Enterprise subscribers get it right away, while users on the free and Go plans get it a day later. Paid tiers run GPT-6 Sol, while the free and Go tiers run GPT-6 Luna. The new feature is “Intelligent UI”: depending on the question, the model responds not just with text but with images, charts, buttons, forms and small interactive tools. ChatGPT can now also start answering while it is still thinking, and according to the company, for questions that need web search, GPT-6 Instant starts answering 44% sooner on average than its predecessor. The models behind Work and Codex are not changing. Source: OpenAI, The Verge.
Anthropic: Claude Haiku 5.5 is out — 75% cheaper on average than Haiku 4.5. The new small model is designed for high-volume, cost-sensitive work (summarization, classification, database queries, customer support, browser use), and also works well as a “sub-agent” alongside the larger Opus 5.5 and Sonnet 5.5 for coding. It is the first Haiku where you can set how much effort (and how many tokens) it spends on a task. For requests under 100,000 tokens it costs $0.10 per 1 million input tokens and $0.50 per 1 million output tokens (compared with $1 and $5 for Haiku 4.5), and $0.50 and $2.50 above that — putting it at exactly the same price level as OpenAI’s GPT-6 Luna. According to developer Simon Willison, the new tokenizer counts around 1.25 times as many tokens for the same text, which amounts to a hidden price increase. Anthropic also halved Sonnet 5.5’s cache read price and is giving Max and Team subscribers monthly API credits (Max 5x: $100, Max 20x: $200). Source: Anthropic, Claude Help Center, Simon Willison.
Microsoft and Nvidia: Windows built for AI agents, and new machines that run local models. At its San Francisco event, Microsoft unveiled Copilot’s “Hybrid Intelligence” features: with your permission, Copilot can also access the files on your PC, take actions in the system and, where it makes sense, use a local model instead of the cloud — the features are coming over the next few months. Windows 11 is getting an isolated runtime environment for AI agents called “Execution Containers,” for all users. The MAI-Code-1.1 Flash coding model (137 billion parameters) runs locally in GitHub Copilot at no extra charge. Available for pre-order are the Surface Laptop Ultra, built on Nvidia’s RTX Spark chip (from about $2,600, shipping October 16), and the Surface RTX Spark Dev Box with 128 GB of unified memory, aimed at models with more than 120 billion parameters ($5,999, from November). Source: The Verge, TechCrunch, NVIDIA, Satya Nadella (X).
SpaceXAI: Grok Bot will now also use other companies’ models. Elon Musk announced on X that Grok Bot will from now on always be powered by the best model for the task at hand, including competitors’ models — citing Anthropic’s Claude Opus 5.5, Midjourney and Suno as examples. According to Musk, simple requests will go to small, fast models and complex ones to large models, and most requests will be handled by a lightning-fast version of the upcoming Grok 4.8. Meanwhile, Grok 4.7 has also become available on Microsoft’s Foundry cloud platform. Source: Elon Musk (X), Elon Musk (X), SpaceXAI (X).
ChatGPT for Teens: OpenAI is satisfied, Common Sense Media calls it an “unacceptable risk.” According to OpenAI’s first report, teenagers use the under-18 version of ChatGPT for less than 15 minutes a day on average, and fewer than 2% spend more than three hours with it in one sitting; the company also announced College Planner, which helps with college admissions, as well as flashcards and quizzes. On the same day, Common Sense Media’s youth AI safety institute rated the service an “unacceptable risk” after more than 4,000 tests: some safeguards (such as refusing sexual role-play) worked, but alerts to parents often failed to trigger in conversations about self-harm. According to the institute, OpenAI should only allow ChatGPT for users over 18 until the problems are fixed. Source: OpenAI, Common Sense Media, BBC.
Letter from fired OpenAI researchers: don’t reduce how well models’ “reasoning” can be monitored. The three safety researchers fired last week, featured in an earlier roundup — Tomek Korbak, Mikita Balesni and Jasmine Wang — wrote a letter to OpenAI’s board and safety committees. According to the letter, as quoted by The Wall Street Journal, the industry does not yet know how to safely develop and deploy models whose workings cannot be monitored, so OpenAI and the other frontier companies should not move forward with developments that further degrade the monitorability of models’ reasoning (their “chain of thought”). They also write that their firing is having a chilling effect on those who remain at the company. According to Gizmodo, OpenAI said it agrees with their recommendations. Source: The Wall Street Journal, Gizmodo.
Google: anyone can now check whether an image, video or audio clip was made with AI. Google has made the SynthID Detector website available to everyone; until now, only journalists and researchers could test it. Besides content from Google’s own models (Nano Banana, Veo, Lyria), the tool also recognizes SynthID-watermarked content from OpenAI, Nvidia and Kakao, with Apple support coming soon; it handles a total of 20 file formats across images, video and audio. Google says it already receives a million verification requests a day, and the feature is also available in the Gemini app and in Chrome. Microsoft and Meta use their own watermarking standards. Source: Google DeepMind (X), TechCrunch.
Biohub: a $1.8 billion partnership for AI models that predict how cells behave. The program is coordinated by Biohub, the nonprofit backed by Mark Zuckerberg and Priscilla Chan: Meta, Google DeepMind and Isomorphic Labs are together contributing $300 million, the US Department of Energy will spend more than $500 million over five years on lab measurements and computing capacity, and the National Institutes of Health (NIH) are contributing datasets produced with earlier federal funding, which Biohub will standardize into a format suitable for AI training. The companies get exclusive use of the data they fund for one year, after which it becomes public; the first dataset is expected in about a year. According to Demis Hassabis, the most important thing AI can be applied to is medicine and human health. Source: Biohub, Reuters, The Decoder.
Ecosia is replacing Mistral with open models, including Chinese ones. The Berlin-based search engine, used by a range of institutions from the German environment ministry to the UK’s NHS, dropped OpenAI for France’s Mistral in May, but is now switching to open-source models. Founder and CEO Christian Kroll told Politico he was disappointed with Mistral’s quality, saying its models are “a year behind,” and that overloaded servers were a frequent problem. Mistral’s chief science officer, Guillaume Lample, encouraged them to try the company’s new model, even through early access. Source: Politico.
Money for AI: Broadcom financing for OpenAI’s chip, $90 million for Nous Research. According to The Wall Street Journal’s sources, Broadcom is arranging more than $50 billion in debt financing for the custom AI chip it is developing jointly with OpenAI, with Apollo and Blackstone among the lenders in talks; the deal is not yet final. Nous Research, which develops the open-source Hermes agent, closed a $90 million Series B round (at a $1.5 billion valuation, according to TechCrunch); it will use the money to build an enterprise version that lets companies run customized agents while protecting their own data. Source: The Wall Street Journal, Nous Research, TechCrunch.
What this means for you: In a single day, the small-model field got cheaper and stronger: Claude Haiku 5.5 and GPT-6 Luna now compete at the same price, so it is worth recalculating what your high-volume AI tasks cost — but check the change in token counting too, not just the list price. On our pricing page we show what you can expect. Meanwhile, Windows, Grok Bot and Hermes all point to AI agents working on your own machine and your own data, and the debates over teens and chain-of-thought monitoring show that this requires clear rules and oversight; if you want to roll out agents, see how we work.