Meta says its internal testing shows comparable performance to GPT-6 Astra, with a 1-million-token context window. Third-party AA-Omniscience results show weaker factual knowledge.
AI news
11 posts
Microsoft announced its Maia 300 custom chip, scheduled for production rollout in 2027. The company plans to scale the Azure Maia family to multi-gigawatt deployments starting in 2027 to serve as its primary AI inference system.
AI researcher Andrej Karpathy estimates that roughly 5,000 people globally have access to internal frontier-grade AI systems. He notes that the outside world has seen a preview of what AI can achieve, while human ability to comprehend its output is starting to fall behind.
TypeSafe AI, the developer of Jev, raised $870 million in a funding round led by Andreessen Horowitz at a $7.5 billion valuation. Jev outputs probabilities instead of text, and the company claims it works significantly faster and uses far fewer tokens than LLMs.
Anthropic released Claude Haiku 5.5, describing it as its cheapest and most capable small model yet, and its fastest at standard speed. According to Anthropic, average running costs are around 75% below Haiku 4.5, factoring in the pricing tiers and an updated tokenizer.
NVIDIA and Microsoft are co-engineering hardware and software to run AI agents locally on Windows. The new RTX Spark-powered Surface Laptop Ultra starts at $2,599 and begins shipping on October 16.
OpenAI is rolling out GPT-6 alongside its interactive Intelligent UI system to over 1.2 billion weekly users. The update launches first for paid tiers, with Free and Go plans following the next day.
Mistral AI has unveiled Mistral Large 4, a 1-trillion parameter multimodal model trained in Europe on 3,800 NVIDIA Grace Blackwell GPUs. A preview API is available now on Mistral Studio, with open weights scheduled for release at the end of the month.
OpenAI released over 700 manuscripts containing AI-generated mathematical results, though only around 42 percent have been formalized in Lean. While researchers noted progress on major problems, several papers have already required revisions or retraction due to errors.
Google has announced Gemini 4 Argon, a frontier model focused on software engineering, enterprise tasks, and cybersecurity defense. According to Google, it features an expanded 1-million output token limit and achieves a state-of-the-art score of 77.9% on DeepSWE v1.1.
Higgsfield claims its annualized revenue run rate has surpassed $1 billion 18 months after launching its AI creative suite. The figure is calculated by multiplying revenue from the latest four weeks by 13, rather than trailing annual revenue.
