Trust & transparency
Evidence & disclosures
Evidence basis: Researched analysis. The article compares public evidence and does not claim hands-on testing unless a section says otherwise.
News roundup based on linked announcements and research.
Sources checked: Oct 16, 2026
Linked references: 12 explicit sources in the article.
Synthetic media: Thumbnail, hero, and social artwork are AI-generated editorial illustrations. Inline artwork and embedded demos are credited to their publishers.
AI had a busy week: more models you can download, a new voice assistant, and a robot with surprisingly good aim. Here’s the October 9–15 roundup, with official demos to watch and links for the details.
Coverage: October 9–15, 2026. Earlier October 9–12 announcements were covered in our catch-up edition; this issue concentrates on developments not covered there.
New Models
Gemini 4 Argon opens its doors
October 13 · Available to the stated paid audiences
Google expanded access to Gemini 4 Argon through its paid API and Google AI Pro and Ultra subscriptions in the Gemini app and Search’s AI Mode. Its official demos range from maintaining Chromium code to building elaborate 3D scenes. This is the wider rollout of September’s announcement, rather than another brand-new model name.
Read Google’s release details →
Watch: Applying Gemini 4 for large-scale codebase maintenance in Chromium — Google DeepMind, official demo →
Meta puts Muse Spark Open on the download shelf
October 15 · Weights released
Meta released Muse Spark Open for coding and multi-step agent work, with downloadable weights and hosted API options. Its Apple Silicon showcase uses two M5 Ultra Mac Studios—not an ordinary laptop. The model has 816 billion total parameters, with 42 billion active at a time. Its Muse Model License includes commercial limits.
Read Meta’s announcement →
NVIDIA brings Nemotron closer to your desk
October 15 · Two models available; one still coming
NVIDIA released Nemotron 3.5 Super VL for agents that read images and documents, plus Bolt 74B-A9B, designed to fit into 24GB of memory for local workloads. A smaller 4GB Bolt E4B is due October 23. NVIDIA’s speed and accuracy comparisons are company claims, so the release details matter more than the victory lap.
Explore NVIDIA’s models and downloads →
Grok Voice 3.0 joins the conversation
October 14 · API available
Grok Voice 3.0 arrived in the API at the same price as its predecessor. The announcement emphasizes quicker spoken responses, interruptions, and customer-support conversations. It cites Artificial Analysis testing for its speech ranking; that doesn’t make it the best assistant for every caller. Developers get a new voice option without waiting for a future launch.
Read the Grok Voice announcement →
Technology Advances
A robot learns cornhole from video. Yes, really.
October 13 · Research demonstration
Google showed Gemini 4 Argon turning a video of a person playing cornhole into simulation code and robot movements. It reports that Apptronik’s Apollo landed seven consecutive shots in a real-world trial. It’s an entertaining example of learning a physical task from video, rather than evidence that general-purpose household robots have arrived.
Read the robotics explanation →
Watch the companion video on YouTube
Watch: Teaching a humanoid robot to play cornhole with Gemini 4 Argon — Google DeepMind, official demo →
From a prompt to a pagoda
October 13 · Official creative demo
Another Google demo shows Argon using Blender and repeated visual critiques to create a detailed pagoda complex. It’s a useful watch for anyone following AI-assisted creative tools: the result is a constructed 3D scene, rather than just a generated still image. Google’s selected showcase isn’t a guarantee that every prompt will produce the same polish.
See Google’s demo collection →
Watch: Pagoda demo: Gemini 4 Argon — Google, official demo →
Anthropic investigates AI running across small devices
October 12 · Controlled security research
Anthropic demonstrated a model running across 15 small TV boxes and studied whether models could set up inference and replicate across a network. The experiments used simplified conditions and identified important limitations. The news is a research result about a potential security problem, not a report of an autonomous AI worm spreading in the wild.
Read the research and its limitations →
Companies Making News
Microsoft reveals the foundation beneath its next models
October 15 · Pre-training complete; post-training underway
Microsoft detailed MAI-Base-2, a foundation model with 2.8 trillion total parameters and 80 billion active parameters. It says training used 16,000 GB200 GPUs over 38 days. The model is still undergoing post-training, so this is a development update—not a new chatbot you can try today. Microsoft’s comparisons use its own evaluations.
Read Microsoft’s lab notes →
Anthropic lets independent evaluators further inside
October 15 · Evaluation arrangement announced
Anthropic says six METR evaluators now have office space, company hardware, and access roughly comparable to its risk teams. METR can publish assessments without Anthropic vetoing publication, though predefined redactions remain possible. It’s a change in how outside scrutiny works; the eventual published findings will be the part to follow.
Read the evaluator-access announcement →
OpenAI offers sponsored access for academic labs
October 15 · Applications and approval required
OpenAI introduced the Rosalind Academic Plan for eligible U.S. university labs, offering up to 30 Business Standard seats at no cost for up to 12 months. Access to its life-sciences model, GPT-Rosalind, requires trusted-access approval. API access is not included. The headline here is expanded research access, rather than a general ChatGPT upgrade.
Check the plan and eligibility →