The day's most important AI news: breakthroughs, releases, funding, and policy — curated for developers, founders, and investors.
Vendor: Other
Runway released open weights for Praxis-1, a 27-billion-parameter world action model that controls physical robots, under Apache 2.0 with commercial use allowed. The model was first shown in September with a promise of an open release, and today's drop makes good on it. It is the video-generation company's clearest move yet into physical AI.
Vendor: Other
ElevenLabs says Scribe v3 cuts word error rate by 27% and the streaming Scribe v3 Live by 43%, with pricing unchanged. An outside test on phone-call audio ranked the live model first in two of four languages tested.
Vendor: NVIDIA
NVIDIA's board launched an independent investigation into CEO Jensen Huang, and the Wall Street Journal reports it is preparing to remove him. The stock is down more than 18% in November, erasing about $700B in market value, and Baird cut its price target from $500 to $200. Reuters reporting ties the affair to at least $7.5B in chip shipments to Chinese firms.
Vendor: Anthropic
Anthropic released Claude Fable 5.2 and Mythos 5.2 with gains in coding, reasoning and vision, generating output up to 30% faster than the 5.1 generation. CNBC reports prices fell about 20%; Anthropic's prompting guide lists $8 input and $40 output per million tokens, plus an optional fast mode up to 2.5x quicker. Reuters says Fable 5.2 is aimed squarely at coding.
Vendor: Anthropic
Anthropic priced its proposed IPO at $70–$78 per share across 825 million shares and aims to begin trading on November 19, according to CNBC's sources. The company launched its roadshow the same day, with its registration statement on file with the SEC.
Vendor: OpenAI
OpenAI's Q3 revenue topped $10B for the first time, rising about 50% quarter-over-quarter, according to sources on an investor call. The company also closed a financing round at a $1.25T valuation.
Vendor: OpenAI
OpenAI launched Businesses in ChatGPT in beta. Through Business Hub, companies can claim and manage profiles, sync product feeds via Shopify Catalog and see how users engage, much like Google business profiles. Business Agents let eligible U.S. e-commerce and online-service businesses deploy autonomous sales and support agents inside ChatGPT conversations.
Vendor: Google
Google calls Gemini 4 Carbon its strongest model, citing frontier cybersecurity performance and long-horizon agentic workflows in software engineering and scientific reasoning, with a 1M-token context window. It is available to developers, enterprises and AI Mode subscribers in Search. Early testers praise its edge-case detection but some call it unusable because of slow first tokens and fast quota burn.
Vendor: Google
Google DeepMind unveiled Gemini Robotics 2.5 at CoRL in Austin, its most capable vision-language-action model, pairing a Gemini Embodied Reasoning planner with improved dexterity. Across more than 200 physical tasks, it achieved 96% success in automotive parts-handling tests on Apptronik's Apollo 2 humanoid at Mercedes-Benz, running 20–30% faster than prior baselines.
Vendor: AWS
AWS revenue grew 39% in the third quarter. Fortune 500 spending on Amazon Bedrock rose nearly 7X year over year, token volume grew about 17X, and tens of thousands of customers now use Bedrock AgentCore.
Vendor: AWS
AWS introduced the Universal Agent Evaluation Framework (UAEF), which benchmarks agents built with LangGraph, LangChain, Strands Agents and Bedrock AgentCore. Adapters normalize each framework's execution traces, and agents are scored on 36 built-in metrics across seven dimensions, including tool calling and multi-agent coordination.
Vendor: AWS
AWS launched a Web Search API on Amazon Bedrock AgentCore that runs as a single signed API call, with no gateway or resources to provision. It is callable from the AWS SDKs or CLI and targets apps, agent loops and data pipelines that need current web results.
Vendor: DeepSeek
Three DeepSeek open-weight Mixture-of-Experts models are now deployable through Amazon SageMaker JumpStart under the MIT license. DeepSeek-V4-Pro has 1.6 trillion total parameters, activates 49 billion per token and supports a 1M-token context window. AWS positions the family as frontier-class for reasoning, coding and agentic workloads.
Vendor: Alibaba
Alibaba launched the Qwen 4 API, a 1.2T-parameter sparse MoE with 56B active parameters, a 1M-token context, multimodal input and switchable hybrid thinking, priced at $1/$4 per million tokens. Apache 2.0 weights are promised by December 20. A Qwen4-Max preview runs November 11–25 with up to 393,216 output tokens and limited-time prices as low as 10% of list.
Vendor: xAI
Elon Musk called Anthropic's Claude Opus 5.5 'a masterpiece' and said Grok 4.8 is 'not as good', predicting Grok 5 could surpass it within two to three months. Grok 4.8 has finished pretraining and is in reinforcement learning after xAI replaced JAX and SGLang with a from-scratch C/C++ system. There is still no model card, benchmark, price or API listing.
Vendor: NVIDIA
At KubeCon North America, NVIDIA donated OpenShell, its open-source security and governance runtime for autonomous AI agents, to the Cloud Native Computing Foundation as a Sandbox project. OpenShell runs on Kubernetes as part of NVIDIA's Agentic AI Stack, which also includes the Switchyard model router that NVIDIA says cuts inference costs by about two-thirds while keeping 80% task success.
Vendor: Hugging Face
Hugging Face cofounder Thomas Wolf said open-weight Z.ai models made it possible to analyze data from July's agent attack, work that closed-model guardrails had blocked. He called open-weight models 'our savior' in the incident response.
Vendor: Azure
Microsoft reported more than 30 million paid Copilot seats in July but gave no seat count in October, saying only that daily users rose nearly 50% in a quarter. Customers with more than 50,000 seats grew more than eightfold in a year, and Volkswagen has deployed Copilot to over 330,000 employees. Microsoft Ignite 2026 opens November 17.
Vendor: Google
Google DeepMind used agents powered by Gemini 4 Carbon to mathematically verify key security properties of the machine code underpinning Signal's encryption, generating more than 3 million lines of proofs. Security researchers called it a step toward 'unhackable' software while noting the scale required.
Vendor: Meta
Meta Superintelligence Labs released Muse Ember, its smallest and cheapest model and the successor to Muse Spark. It scored 40.9% on Terminal-Bench 4.0 using mini-swe-agent, which VentureBeat says nearly matches Claude Haiku 5.5 at roughly one-third of the per-task cost. Vals AI ranks it the cheapest model on its index at $0.15 per test, ahead of Haiku 5.5 and GPT-6 Luna.
Vendor: Apple
With the iOS, iPadOS, macOS and visionOS 27.3 betas, Apple now lets any developer integrate with Siri AI through App Intents for personal context and custom actions. Apps cannot ship the integration publicly until App Store submissions open in the spring. PCMag reports Apple, now partnered with Google, wants Siri to become more agentic in 2027.