AI Newsletter

OpenAI officially launched GPT-6 Astra on September 3rd, calling it the 'world's most intelligent and most aligned model' with near-perfect scores in coding, cybersecurity, and scientific reasoning. The company states that the model marks the beginning of the AGI era – however, internally the safety team expresses serious concerns about the observability of chain-of-thought. Despite the launch, Polymarket (86%) continues to show Anthropic as the favorite for the best model by late September 2026, suggesting an imminent response from Anthropic.

Thomson Reuters has developed a proprietary LLM called 'Thomson' based on Qwen3.5-397B, trained over two years for approximately 40 million dollars and containing 175 years of legal, tax, and accounting data. The move symbolizes a broader enterprise trend: major customers are actively cutting spending on OpenAI and Anthropic in favor of cheaper or self-developed alternatives. This increases pricing pressure on frontier labs significantly and could structurally threaten the API business model of major AI providers.

Anthropic communicates to investors a potential revenue opportunity exceeding 30 billion dollars and is preparing for an IPO, to which Polymarket assigns a 96% probability before September 30th, 2026. In parallel, Anthropic dominates model rankings and has just released Fable 5.1, a more cost-effective update that according to initial analysis significantly improves the pricing structure compared to Fable 5. The IPO would revalue the AI sector and could accelerate OpenAI's own IPO ambitions (63% chance of $1.5T market cap by end of 2027).

21 top researchers from Stanford, Oxford, DeepMind, CMU, and Meta have published a paper ('Visual Generative...') arguing that LLMs are a dead end on the path to AGI. The criticism arrives precisely at the GPT-6 Astra launch, suddenly placing scientific disagreement about the correct AGI path at the center of attention. For companies building their AI strategy on LLM APIs, this represents strategic uncertainty – particularly if alternative architectural approaches gain momentum.

Nvidia has signed a definitive agreement to acquire Hugging Face for 12.9 billion dollars, with planned closing in the first half of 2027. The acquisition would give Nvidia direct access to the world's most important open-source model repository for the first time and complete vertical integration from hardware through infrastructure to models. This fundamentally changes the power dynamics in the AI ecosystem: whoever hosts on Hugging Face or sources models would in the future be indirectly dependent on Nvidia's infrastructure.