Weekly AI #1: Gemini 4 Argon, faster Claude and open models everywhere

Claude now starts answering up to 3x sooner

Anthropic has cut the wait before Claude starts answering. Between August and October 2026, the time to first token dropped about 3x in Claude Code and about 1.5x in Claude chat, Cowork and the API, by Anthropic's own measurements.

Anthropic

Gemini 4 Argon controls robots and builds simulations to train them

Google has shown Gemini 4 Argon controlling real robots and building physics simulations to rehearse in. On ERQA, a test of reasoning about the physical world, Argon scores 87% with no tools, which Google says beats every other frontier model.

Google

Gemini 4 Argon reaches Google AI Pro in Antigravity with a weekly quota

Google is rolling out Gemini 4 Argon in Antigravity to paid Google AI Pro subscribers. The Antigravity team announced it on X on October 15, 2026. Pro users get a separate, weekly-only quota for Argon on top of their existing Gemini 3.8 Flash quota.

Google

DeepSeek V4.1-Flash-1017: higher scores, shorter answers, new Search API

DeepSeek has released DeepSeek-V4.1-Flash-1017, an update to V4.1-Flash. DeepSeek says it scores higher, finishes tasks faster and writes shorter answers, which makes it cheaper to run. The weights are open under the MIT license, and DeepSeek says the model is on its website, in…

NVIDIA Nemotron 3.5 Super VL: an open multimodal model for always-on agents

A 120B open-weights model that reads text, images, documents and two-minute videos, plus Bolt, a compressed 74B version that fits in 24 GB. How they are built, how they score and where to run them.

Has Brains AI

Black Forest Labs releases FLUX 3 Image Klein, a 5.6B open-weights model

Black Forest Labs has released FLUX 3 Image Klein, a 5.6B-parameter open-weights model for text-to-image generation, image editing and multi-reference generation. Base and distilled weights are on Hugging Face under the FLUX Kommunity License. The distilled model is also in the …

Google's Harness² lets an AI agent tune its own harness

Google Research has released Harness², an open-source framework that lets an AI agent improve its own harness for a new task without retraining the model. The harness is everything around the model: system prompt, tools, memory, skills, subagents and more. The code is on GitHub …

Google

Perplexity open-sources SPACE, the sandbox platform behind Computer

Perplexity has open-sourced SPACE, the sandbox platform that runs code for its Computer agent. The code is on GitHub under the Apache 2.0 license, with hosted documentation. SPACE stands for Sandboxed Platform for Agentic Code Execution. It gives each agent its own isolated virt…

Are neural networks already clean? Nick Cammarata's case against superposition

An interpretability researcher argues most neurons already do one readable job, and that the famous cat-and-car neuron is really a V-shape detector. What it could mean for AI safety research.

Has Brains AI

OpenAI offers U.S. university labs free GPT-Rosalind access

OpenAI has launched the Rosalind Academic Plan for academic labs at U.S. universities. Eligible labs can get up to 30 ChatGPT Business Standard seats at no cost for up to 12 months. They also get access to GPT-Rosalind, OpenAI's life sciences model, after a trusted-access review…

OpenAI

Anthropic gives six METR staff employee-like access for risk reviews

Anthropic has given six employees of METR, a nonprofit that evaluates frontier AI models, what it calls "employee-like access". The team gets company hardware, office space and access to internal tools and systems, roughly on par with what Anthropic's own risk assessment teams h…

Anthropic

Anthropic opens Singapore office and signs MOU with the government

Anthropic has opened an office in Singapore, its fifth in Asia-Pacific. On October 14, 2026, it also signed a Memorandum of Understanding (MOU) with Singapore's Ministry of Digital Development and Information. The MOU covers AI safety, security and model evaluation.

Anthropic

Manus launches Team and Enterprise plans for businesses

Manus has launched business plans. The Team plan is available now and starts at US$20 per paid seat per month, with 4,000 credits per seat. Enterprise adds stricter governance for large organizations and goes through sales.

Mistral acquires New Material and launches a Solutions Design team

Mistral has acquired New Material, a design-led AI product company. Its team will form a new unit at Mistral called Solutions Design. New Material announced the deal on October 14, 2026. The terms were not disclosed.

Amazon launches Bee Frontier and Bee Eclipse AI devices

Amazon has launched two Bee devices that turn conversations into summaries and to-dos. Bee Frontier is a wearable that costs $74.99 and is available now at Amazon.com. Bee Eclipse is a phone attachment that starts at $99.99 and will be available soon. Some features need Bee Pro,…

OpenAI trains staff to spot spies and honey pots, WSJ reports

OpenAI has been training employees to spot espionage attempts, from romance "honey pots" to strangers asking about their work, The Wall Street Journal reported on October 14, citing unnamed sources. Staff are told to avoid company logos in public, hide their screens and wipe whi…

OpenAI

RUMOR: Nvidia is moving GB202 chips from the RTX 5090 to RTX Pro cards

Leakers claim Nvidia has stopped sending new GB202 chips to the GeForce RTX 5090 and is keeping them for its professional RTX Pro workstation cards. If true, that would explain why the RTX 5090 has become so hard to find.

Meta releases Muse Spark Open, an 816B open-weight model that runs on two Macs

Meta released the weights of Muse Spark Open, an 816B-parameter multimodal model built for coding and agents, which Meta calls "the most capable US open-weight model to date". It runs locally on two Mac Studios with M5 Ultra, and is available today on Hugging Face and through AP…

Anthropic: AI models may soon copy themselves across cheap smart devices

Anthropic's Frontier Red Team says AI models may soon be able to copy themselves across cheap smart devices, like TV boxes and cameras, with no datacenter and no GPUs. A network like that would have no centre, which makes it hard to shut down. Today the risk is low: replication …

Anthropic

US nuclear plants adopt AI assistants for paperwork, not control

Most US nuclear reactors now use AI assistants for paperwork and search, not for running the plant. IEEE Spectrum reports that nearly all 94 US reactors have been offered AI tools, and most have taken them up.

DeepSeek to lead seed round in robotics AI startup Winverse AI

DeepSeek is set to lead a seed round in Winverse AI, a Chinese startup building embodied AI models for robots. Bloomberg reported the deal on October 12, 2026, citing people familiar with the matter. It is not final yet.

Tesla files a patent for touch-sensitive skin for robot hands

Tesla has applied for a US patent on soft, curved touch sensors for robot hands. The drawings show a robot hand with one sensor along the thumb, one covering most of the palm, and one on the index finger. The application was published on October 8, 2026.

PS5LM runs llama.cpp and local LLMs on a jailbroken PlayStation 5

PS5LM, an open-source project on GitHub, ports llama.cpp to the PlayStation 5, so the console runs local language models on its own CPU and GPU. Its author reports Qwen 3.8 27B answering at 9 to 21 tokens per second on the console's GPU. It needs a jailbroken PS5.