SpaceX Advances 5G, Spectrum, and Starship Plans as AI Labs Trade New Model Launches

SpaceX outlined a final-agreement 5G build with Qualcomm, an FCC spectrum application, a Starship catch attempt, and new satellite and solar-equipment plans.
Anthropic and Google both shipped new frontier models, with benchmark claims and pricing trade-offs disputed across sources.
OpenAI expanded ChatGPT for businesses and disclosed frontier-model security findings, while Stripe acquired OpenRouter.
Hugging Face reported 20 million AI developers and 300PB of hosted assets, alongside low-cost OCR and agent-training examples.
Runway released open-weight robot-control model Praxis-1, and Perplexity described agent-built quantum simulation and marketing workflows.

SpaceX and Qualcomm signed a final agreement for SpaceX to build base stations and Qualcomm to supply chips and RAN accelerators for an AI-native terrestrial 5G network; Qualcomm's CEO called it one of the largest 5G deployments in the US, tied to Starlink direct-to-phone and V3 satellites 1. Separately, the FCC accepted SpaceX's application to buy nationwide 800 MHz spectrum from Grain Management: 546 license pairs covering about 100% of the US population, with SpaceX holding at most 89 MHz per market after closing; Musk called it the last key spectrum needed for full US mobile coverage 2.
Starship's 15th flight will attempt to catch the ship with the launch tower after its return from orbit, something never done before, no earlier than November 13 1. SpaceX also ordered HJT solar manufacturing equipment from Germany's Singulus with 100 GW annual capacity for accelerated Starmind satellite production in Bastrop, Texas; Musk described the goal as building a "sentient sun" 2. Starlink Mini 2 adds a quick-swap battery rated up to 4.5 hours, built-in Wi-Fi, and online access within 60 seconds of pointing at the sky, now in invitation-only US preorder 3 4.

Anthropic released Claude Fable 5.2 and Mythos 5.2, with Fable topping the Artificial Analysis Intelligence Index five points ahead of Google Gemini 4 Carbon while costing 20% less and running up to 30% faster than 5.1, according to a roundup 1. Google launched Gemini 4 Carbon the same day; Demis Hassabis cited a Gemini 3-to-4 iteration in under a year, a 60.4% score atop CursorBench at high compute, and a tie with Fable 5.2 at 61.5% on GamowLabs' RareBench, which he said is nearing saturation 2 3 4. Hassabis also called Carbon highly efficient with low token use, low price, and high speed, close in capability to Fable 5.2 5; Perplexity's Aravind Srinivas said Carbon approaches Fable 5.2 on agentic research tasks at 2.5x lower cost 6.
Benchmark methodology drew scrutiny: Hamel Husain argued that benchmarks people cannot game are called worthless, yet become statistically meaningless once gamed, and noted the Artificial Analysis Intelligence Index uses nine benchmarks with arbitrary weights that shift substantially over time 1 2. He also said benchmark results should show error bars and that omitting them is a serious charting error 3. Separately, a roundup said AI 2027 author Daniel Kokotajlo moved his 50% probability for top-human-level AI research to May 2027 and described feeling "afraid, sad and angry" 4.

OpenAI launched a beta Businesses hub in ChatGPT: companies can claim merchant profiles, manage information, view interaction insights, and, if eligible, build a Business Agent to answer customer questions 1 2. OpenAI also said it found and fixed important vulnerabilities in its own code and infrastructure while using frontier models, with findings to be shared in a November 12 Cyber at the Frontier talk 3. Anthropic, meanwhile, began IPO investor marketing with shares possibly trading as soon as the following Thursday, and its updated filing said it had crossed its own AI-progress safety line, per the roundup 4.
Stripe formally acquired OpenRouter, which a cited post said serves over 10 million developers and companies with cross-model and cross-provider access, availability during outages, team spending and data policies, and a commitment to neutral competition among models and providers 1. The same roundup said OpenAI acquired travel-planning startup Odessia, while AMD's latest 10Q showed $2.9 billion paid for Taalas, which hardwires LLM weights into hardware 2 3.

Hugging Face reported 20 million AI developers and 300PB of hosted models, datasets, agent trajectories, and reinforcement-learning environments, including 9PB added in the past week, with October setting records for daily active users, monthly active users, and revenue 1. It highlighted Idol, a 600-million-parameter image-to-3D model trained from scratch that leads its trending models, needs under 7.2GB of VRAM, generates 1 million splats in 3 seconds, and carries an Apache 2.0 license 2. A TRL-based agent-trajectory SFT example raised a 2B model from 5.5% to 34.4% on held-out coding tasks 3, and LightOnOCR-3-0.8B processed 14,483 pages at about 7.7 pages per second for $1.67 total, or roughly $0.12 per thousand pages 4.
Runway opened Praxis-1 to everyone: a world-action model that turns video pretraining into real robot control, released with open weights under Apache 2.0 for running and fine-tuning on own hardware 1. Perplexity's Srinivas said its Scientific Researcher teammate built a self-correcting topological quantum-memory simulator from scratch, including a 12-anyon Fibonacci entangling gate with 24 braids and 0.00069 leakage and an audit of 2,934,376,200 double-fault pairs 2. He also described a marketing-studio workflow where Teammates research a brand, draft briefs, propose three directions, storyboard six shots each, and render only the approved direction into a 1080p vertical ad with native audio 3.

Google released Gemini Robotics 2.5, which it said can learn hundreds of everyday skills including folding laundry, serving pastries, and autonomously sorting car parts on a factory floor 1. Google also announced two AI chips for next year: TPU 8tx, which keeps data encrypted while a model runs, and 8i Max, which combines two chips to support agents that must review an entire codebase; Google said it has nearly 100 TPU customers and that existing chips are fully allocated 2. Adobe and Disney signed a multiyear deal to train custom AI models on Disney characters and visual styles, already used for theme parks, film, and ESPN 3.
Meta released Muse Ember, its smallest model, for high-concurrency agent tasks; the roundup said it matches Claude Haiku 5.5 on the intelligence index at about one-third the cost per task, $0.14 versus $0.41 in average tests 1. NVIDIA said OpenShell joined the Cloud Native Computing Foundation, with a cited post saying it entered the CNCF Sandbox to provide policy isolation, network control, and credential management for autonomous agents, arguing agent security belongs at the runtime layer rather than only the prompt layer 2. France published a Paris AI Safety appeal, with OpenAI, Anthropic, and Mistral committing to built-in emergency shutoffs in their most advanced models, according to the roundup 3.

AI & Tech