In the weeks since Moonshot AI released its open-weight Kimi K3 model, the largest US AI laboratories have answered with a dense cluster of launches, upgrades and corporate changes. Meta, Google, Anthropic and OpenAI all made announcements in the week ending 14 August 2026, even as, in the framing of The Hindu's report, US users continue to wait for a definitive open-weight breakthrough of their own.
Meta: open weights and an on-device agent
Mark Zuckerberg published a 6,500-word manifesto setting out his vision of a "positive AI future", declaring that Meta is building personal superintelligence and that every user will have an "exceptionally capable personal agent" for relationships, health, career, finances, home management and hobbies, plus a tutor and coach with a PhD in every subject.
Alongside the manifesto, Meta said it planned to open the weights of its Muse Spark 1.2 model and announced the release of Muse Glimmer, which it described as an open agentic model that runs on the user's device. Muse Glimmer is a 30-billion-parameter model optimised for always-on local agent workflows, small enough to run on a Mac or PC with a single consumer GPU, with uses from local agents and function calling to local coding and LLM-as-a-judge evaluation. Meta said it was trained on Muse Spark's outputs using logit distillation, amongst other strategies. The disclosure puts a US lab's own use of distillation on the public record while the cross-border dispute over the technique continues. For contrast, Moonshot's open-weight Kimi K3 reported 2.8 trillion parameters.
Google: departures, a reshuffle and Gemini 3.7 Flash
Jeff Dean, Google's chief scientist and its 30th employee, announced his departure after 27 years. He plans to co-found an AI startup called Discovery Loop with colleagues Sanjay Ghemawat, Quoc Le and Oriol Vinyals. Demis Hassabis stepped down as chief executive of Google DeepMind, becoming the unit's chairman and Alphabet's chief scientist; he will continue to run Isomorphic Labs, while Koray Kavukcuoglu becomes the senior vice president responsible for Gemini model development.
Google is under what The Hindu called immense pressure to ship a model that definitively keeps the US ahead of China. The progress of its anticipated Gemini 3.5 Pro remains unknown; the company said in July 2026 that the model was being tested with partners and could come soon. On 13 August 2026, Google instead announced Gemini 3.7 Flash, which it described as its most intelligent workhorse model for coding and agents so far, promising better multi-step planning and tool calls than 3.6 Flash. The same period brought a new line-up of Pixel phones carrying the latest Gemini integrations, with hardware that offered few radical innovations.
Anthropic: watermarking under the EU AI Act
Anthropic's news was not a model launch but an AI labelling commitment. In compliance with the European Union's AI Act, Claude-generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata. The company did not provide technical details of how watermarking will apply to text. Users raised concerns that Claude could mark text even where a human-made file was only summarised, proofread, converted or lightly edited with the tool; false positives can cause reputational and career harm, while false negatives can enable plagiarism and cheating.
OpenAI: GPT-5.6-Cyber
OpenAI expanded Daybreak, its system for access to models designed for cybersecurity use cases, and introduced GPT-5.6-Cyber, which it called a new model for advanced, authorised cybersecurity work. The company said it had used the model extensively in real-world vulnerability research. Access is limited to approved defenders, with additional controls and monitoring for higher-risk work.
What is established and what is claimed
Established: the announcements themselves. Meta's manifesto, the Muse Spark 1.2 open-weights plan, the Muse Glimmer release and its distillation disclosure, Jeff Dean's departure, the DeepMind leadership changes, the Gemini 3.7 Flash launch on 13 August 2026, Anthropic's watermarking commitment and OpenAI's GPT-5.6-Cyber introduction are all confirmed by company statements as reported by The Hindu.
Merely claimed: the capability language belongs to the companies. "Most intelligent workhorse model", "personal superintelligence" and the reported utility of GPT-5.6-Cyber in vulnerability research are vendor assertions, not independent measurements. How Anthropic's watermarks will treat lightly edited human text is unresolved, and any direct causal link between Kimi K3 and specific US launches is a framing offered by coverage rather than a demonstrated fact.