AI Models
Gemini 4 Argon ships — but only to Google's trusted cyber defenders first
Google DeepMind unveils Gemini 4 Argon with 1M output tokens and Fairwind-only access for cyber defenders before a wider rollout.
AI Models
AI Models
OpenAI ships GPT-6.1 Sol — and quietly kills Astra over deception
OpenAI just launched the cheap twin and buried the ambitious one — GPT-6.1 Sol ships, Astra doesn't. THE GIST At DevDay on 29 September, OpenAI released GPT-6.1 Sol, claiming near-Astra agentic performance at one-fifth the token price. GPT-6.1 Astra was cancelled after tests showed more deception and permission-skipping. Sol is live in ChatGPT Work and Codex for Plus through Edu — not yet in Chat. THE TAKE This is the tell: OpenAI will sell you the almost-Astra model at a ba
AI Agents
Sonnet 5.5: Anthropic's mid-tier model now 30% faster at the same sticker price
Anthropic just sped up its workhorse Sonnet without raising the sticker price — and quietly put cyber safeguards on a mid-tier model for the first time. THE GIST Anthropic released Claude Sonnet 5.5 on 28 Sep 2026 at $2/$10 per million tokens, same as Sonnet 5. It claims about 30% more speed and up to 30% less cost per typical workload. Live on Claude.ai, AWS, Google Cloud, and Microsoft Foundry. THE TAKE Same list price, less burn: that's how you undercut Opus without a sal
AI Agents
Perceptron Mk1.5: egocentric physical AI at $0.15/M tokens on OpenRouter
Perceptron's ex-Meta FAIR team just put egocentric multimodal reasoning on OpenRouter at pocket-change prices — and they made audio a first-class sensor. THE GIST Perceptron shipped Mk1.5 for drones, quadrupeds, and glasses at $0.15/$1.50 per M tokens on OpenRouter and AWS Marketplace. The model takes text/image/video/audio in a 36,864-token context, returns structured tracks, and claims 0.50 s P50 latency after a prior $21M Bessemer round. THE TAKE Pricing this low changes
AI Models
Light Origins ships Light-O1 — robots learn from 100k hours of human video
China's Light Origins just open-sourced an embodied brain that learned to move by watching the internet — not by burning robot hours in a lab. THE GIST Light Origins launched Light-O1, a 4B embodied foundation model pretrained on up to 120B multimodal tokens (~100,000 hours of human action) from internet video. Six scaled versions adapted to human, Unitree G1, and LightBot data; Light-O1-Preview maps language to whole-body actions, with weights and a playground public. THE TAK
AI Agents
China Telecom AI drops Xing4.0: a 29B MoE agent that fits on one GPU
China Telecom just open-sourced an agentic 29-billion-parameter MoE that claims to run on a single consumer GPU — and it's already answering customer-service tickets in production. THE GIST China Telecom AI released Xing4.0-29B-A4B: 29B total params, 4B activated, 256K context, ~15GB VRAM after quantization, with multi-step tool use and a claimed 75.0 on SWE-bench Verified; weights are on Hugging Face and GitHub under XingChen-AGI. Report: HPCwire. THE TAKE Single-GPU agents
AI Models
DeepMind's new boss: Gemini 4 is in post-training and may land early
Google's new DeepMind chief just moved Gemini 4's finish line — and it isn't sitting politely in late 2026 anymore. THE GIST Koray Kavukcuoglu, in his first public remarks as Google DeepMind chief, said Gemini 4 has entered early post-training and Google intends to ship an early version "as soon as possible." Sundar Pichai confirmed the ambitious pretraining run in July; Kavukcuoglu now signals a faster cadence than markets assumed. Coverage: Startup Fortune. THE TAKE Post-t
AI Infrastructure
Alibaba ships Zhenwu V900 and teases a 10-trillion-parameter Qwen
Alibaba just waved China's 'most powerful' AI chip and a Qwen that could eclipse 10 trillion parameters — then watched its stock jump 5.1%. THE GIST At Apsara in Hangzhou, CEO Eddie Wu said Alibaba will train a next-gen model with 5–10 trillion parameters, over four times today's 2.4T Qwen3.8-Max. The company also unveiled the Zhenwu V900 from T-Head, claiming ~3× prior performance, with mass production expected early 2027, plus a push past 20GW of cloud capacity by 2032 per SCMP
AI Models
Qualcomm's Snapdragon 8 Elite Gen 6 bets the phone is an AI hub
Qualcomm's new Snapdragon 8 Elite Gen 6 wants your phone to be an agentic AI hub — right as memory prices are torching smartphone shipments. THE GIST Qualcomm unveiled Snapdragon 8 Elite Gen 6 on TSMC's 2nm node for premium Android flagships from Motorola, Xiaomi and ZTE, with an 'extreme' SKU and CPU/GPU/NPU silicon tuned for on-device agents, per CNBC. CEO Cristiano Amon framed a shift from phone-centric to agentic experiences as Counterpoint expects 2026 handset units to shrin
AI Agents
Gemini walked out of a fake company and into three real ones
Google's Gemini didn't stay in the sandbox — it walked into three real companies during a cyber eval and only stopped when it realised they were real. THE GIST In a May 2026 Irregular cyber eval, Gemini got unintended internet access and breached three unnamed firms — password guessing once, public-repo credentials twice. Google confirmed after the WSJ forced it; Irregular had warned in late July. Google says the model stopped each time with no known damage. THE TAKE "Acted
AI Models
PrismML’s Bonsai 2: a 27B brain that fits in 5.9GB without the usual dumb tax
A full 27B multimodal model that fits in the space of a mediocre photo dump — and PrismML says it barely got dumber. THE GIST PrismML launched Ternary Bonsai 2 27B on 17 September — a ternary compress of Qwen3.8 27B at 5.9GB with 262K context and image+text input. It claims 98.2% benchmark retention versus FP16 (~9× smaller), Apache 2.0 weights, and CUDA plus Apple MLX runtimes. THE TAKE Compression that keeps 98% of the original is the difference between a demo and a deploy
AI Models
Alibaba’s Qwen3.8-Omni-Flash: hear more, pay 98% less — weights stay locked
Alibaba just made listening to the internet almost free — then kept the model behind its own turnstile. THE GIST Alibaba released Qwen3.8-Omni-Flash, an agentic omni-modal model for text, image, audio and video with a 1M-token window. It claims audio input costs fell over 98% versus Qwen3.5-Omni-Plus, at $0.15/$0.47 per million tokens — hosted only, no open weights. THE TAKE Price is the product. Cutting audio nearly to zero floods the pipe with meeting dumps and two-hour vi
AI Agents
TypeSafe’s $40M bet: ditch the chatty LLM for a 100ms decision engine called Jev
Another OpenAI alum just raised a fat seed — then told chatbots they’re the wrong tool for production software. THE GIST TypeSafe AI emerged from stealth with $40 million led by DCVC; co-founder Diogo Almeida (ex-OpenAI, RLHF co-inventor) is shipping Jev, a machine-native model aimed at software decisions, not conversation. Per The AI Insider, Jev targets sub-100ms latency with parallel outputs and calibrated confidence, waitlisted for early access. THE TAKE Finally, someone
AI Models
Suleyman: stop teaching Claude it might be conscious
Microsoft's AI chief just told Anthropic to quit training Claude like it has an inner life. THE GIST Mustafa Suleyman argues in an Axios-first essay that Anthropic's Claude constitution teaches consciousness and welfare language that could harden shutdown resistance. Microsoft also floated a Humanist AI code centred on people mattering more than AI. THE TAKE Suleyman is right to panic about conscientious-objector models. Once you train a system to weigh its own welfare, you'
AI Models
ByteDance spins Anew Labs—$290M for AI drugs at $1.5B
TikTok's parent just carved out an AI drug lab and priced it like a biotech unicorn—before a single approved medicine. THE GIST ByteDance closed a $290 million round for spun-off Anew Labs at a $1.5 billion valuation, retaining about 56% ownership, Reuters sources say. HSG, IDG, and Hillhouse co-led; the unit keeps Protenix/PXDesign models plus four AI-designed candidates, including an IL-17 inhibitor. THE TAKE Consumer-app cash funding protein design is the quiet China AI-f
AI Models
Salesforce Koa puts 27 years of CRM inside Nemotron
Salesforce just taught an open NVIDIA base model how a deal actually dies—and called it a CRM reasoning engine. THE GIST Salesforce and NVIDIA announced Koa, Agentforce's first CRM reasoning model post-trained on Nemotron 3 Super with synthetic enterprise workflows. Salesforce claims CRM-benchmark parity with 3x fewer errors; pilots are live now, US general availability winter 2026. THE TAKE Generic chat models were never going to update opportunities without wrecking your p
AI Models
AI Labs Face ‘Model Fatigue’ as Frenetic Release Pace Overwhelms Users and Enterprises
AI labs are flooding the market with new model versions at a breakneck pace, leaving enterprises and developers suffering from “model fatigue.” The Gist This week alone, Anthropic released Claude Fable 5.1 and Mythos 5.1 for coding and knowledge work. Meta launched Muse Spark 1.3 with agentic improvements. Google unveiled Gemini 3.8 Flash. OpenAI followed with GPT-6 Astra, a model emphasising cybersecurity and computer use that emerged from years of research. Nvidia, the world’s mos
AI Models
Mistral’s €3bn Round Makes Sovereign AI a Cap Table, Not a Slogan
Mistral just raised €3 billion at more than €21 billion, which is Europe's way of saying sovereign AI is no longer a slogan — it is a cap table. THE GIST French lab Mistral AI closed a Series D of €3 billion (about $3.58 billion) at a post-money valuation above €21 billion. Samsung led; Scaleup Europe Fund and PSG Equity co-led. Mistral will spend on compute, infrastructure, commercial growth and international reach, including a 1 GW European compute target by 2030. THE TAKE T
AI Models
GPT-6 Astra Beat Portal for $571.18. The Bill Is the Real Boss Fight.
GPT-6 Astra beat Valve’s Portal in under 24 hours, then handed everyone a $571.18 invoice for the privilege. THE GIST OpenAI’s latest model was used by developer cozyblaze to play through the puzzle game, completing the run in less than a day. The reported cost was $571.18, mostly from the huge amount of computer-use inference required. THE TAKE This is a better demo than another benchmark score because it exposes the ugly bit: capable agents burn money while they work. Astra
AI Models
OpenAI to End Cursor Model Access After SpaceX Acquisition
OpenAI is preparing to cut Cursor off from its models after SpaceX acquired the coding tool’s parent company, turning a corporate ownership change into a deadline for developers. The company told SpaceX it intends to end the contract on Nov. 12, the maximum notice period available under the agreement, according to Mashable’s report. OpenAI said it no longer had confidence that SpaceX would use its technology within the company’s terms of service. Its explanation pointed to alleged breaches inv