AI News Summary 2026-09-25

Miguel Fiandor

2026-09-25 14:39

Also available in:

AI News Summary — 2026-09-25

A daily read on current developments in artificial intelligence and related technologies across major AI companies, technical blogs, generative imagery, agents, and local serving.

1. Major AI companies and GAFAM

Anthropic has signed a strategic agreement with xAI and SpaceX to use computing capacity from the Colossus II supercomputer, reportedly worth $1.25 billion per month through May 2029. In parallel, Google has launched official support for the Gemini 3.7 Flash model family, with advanced reasoning and server-side tool execution capabilities Simon Willison's Weblog on Anthropic.

2. Influencers and technical blogs

OpenAI has announced the limited rollout of GPT-6 Astra for Plus, Pro, Enterprise, API, and AWS users, reportedly reaching 99.9% accuracy on the ARC-AGI 3 benchmark through a custom adapter harness that preserves reasoning state between calls. Hugging Face has completed a reevaluation of 3,751 historical models on its Open LLM Leaderboard using the Math-Verify library Simon Willison's Weblog.

3. Generative imagery

Google has released the open-weight google/diffusiongemma-26B-A4B-it diffusion model under the Apache 2.0 license. NVIDIA has incorporated it into its NIM cloud infrastructure, with reported speeds above 500 tokens per second. Pencil AI has also integrated support for the new GPT-5.6 models into its governed generative design platform Simon Willison on Google.

4. Chatbots and agents

Nous Research has released Hermes Agent v0.20.5, consolidating stability improvements for multi-agent environments. OpenJarvis v1.0 has also launched as a local-first framework for deploying personal autonomous agents with native Ollama support Ollama Blog on OpenJarvis.

5. Local AI and serving

Ollama has announced new Pro, Max, and Team plans based on transparent token pricing with included monthly credits and no fixed time limits. vLLM and Unsloth have released updates focused on optimizing GPU memory management for high-speed local inference workflows Ollama Blog Pricing.