Last week compressed a generation-level model launch, the largest platform acquisition in the open-source AI ecosystem, and a congressional attempt to criminalize superintelligence into approximately 72 hours. Google, operating on a different cadence, spent August reshaping the cost structure underneath.
The new generation
OpenAI launched GPT-6 Astra on September 3, the first full generation number since GPT-5. Astra operates software through screens, and OpenAI has demonstrated it filling forms, updating CRM records, editing spreadsheets, and driving engineering tools such as KiCad and FreeCAD. OpenAI reports that Astra scores 72.6 percent on OSWorld 2.0, a benchmark for operating desktop software, against 65.7 percent for GPT-5.6 Sol. The model carries a one-million-token context window and costs $10 per million input tokens and $50 per million output, 2.5 times Sol’s pricing, positioning the generation jump as a premium capability tier above the existing lineup.
The model introduced recurrent depth, a reasoning technique that partially obscures the chain of thought, drawing concern over whether external monitors can verify how the model reaches its answers. Astra also became the first OpenAI model to reach the Critical cybersecurity threshold under the company’s Preparedness Framework. OpenAI restricted access to the most sensitive capabilities, gating them behind a trusted-access program called Daybreak while paid ChatGPT tiers and the API followed over subsequent days. OpenAI president Greg Brockman described Astra as the arrival of artificial general intelligence, though independent benchmarks show mixed results against Anthropic’s frontier models, with Claude Fable 5.1 outperforming Astra on Humanity’s Last Exam.
Buying the marketplace
Nvidia confirmed on September 3 that it would acquire Hugging Face for $12.9 billion, its second-largest acquisition after the $20 billion purchase of Groq assets in December. The deal gives Nvidia control of a platform hosting three million models, one million applications, and eighteen million developers on annualized revenue of roughly $150 million, a valuation premium driven almost entirely by strategic position. Nvidia has committed to keeping the platform open and supporting other silicon vendors. Hugging Face CEO Clem Delangue told CNBC that he approached Jensen Huang directly, describing Nvidia as “a perfect home,” and that discussions moved quickly once they began.
The acquisition positions Nvidia across the full AI stack, from chip design through the developer ecosystem in which models are shared and deployed. Anthropic and OpenAI are both developing proprietary inference chips to reduce their dependence on Nvidia hardware, and ownership of the open-source platform hedges against the possibility that the company’s largest customers become its competitors. Between the Groq and Hugging Face deals, Nvidia has spent more than $30 billion on acquisitions in under twelve months, a pace that reflects confidence that controlling the ecosystem around its chips matters as much as the chips themselves. The Hugging Face deal requires regulatory approval and should close in the first half of 2027.
The cost curve
Google shipped Gemini 3.7 Flash on August 13, three weeks after 3.6 Flash went stable, at an introductory price equal to half the original 3.6 Flash cost per million tokens. The model targets coding and agentic workflows, and Google has positioned aggressive cost reduction as its competitive strategy while other labs spent the same period competing on capability milestones.