AI News: October 24, 2026

1. Moonshot’s Kimi K3.1 Pro Reaches Fourth on the Artificial Analysis Intelligence Index

Moonshot AI. Moonshot released Kimi K3.1 Pro, a post-trained update to Kimi K3, plus a K3.1 Pro Highspeed variant it says runs two to three times faster. K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index (up 11 points from K3, behind Claude Opus 5.5 at 58), 65.7% on Terminal-Bench 4.0 (versus 66.4% for Opus 5.5), and cuts its internal hallucination rate from 23.4% to 8.3%. Pricing is $2 input and $10 output per million tokens, about 33% below K3, and Moonshot plans to release the weights in early November, which would make it the highest-scoring open-weights model. Source

2. UK Commits £100 Million to an AI Biosecurity Lab, With Up to £200 Million From OpenAI and Anthropic

UK Government. The UK government announced the AI Biosecurity Lab, an independent non-profit it calls the world’s first dedicated AI biosecurity organisation, focused on early detection of biological threats and faster development of vaccines and countermeasures. The government is putting in £100 million, while OpenAI and Anthropic are each committing up to £100 million in compute, model credits, and trusted model access. Google DeepMind and Basecamp Research join as research partners, alongside the UK Health Security Agency and other public bodies. Source

3. OpenAI Lowers Its Valuation Target to $1.25 Trillion for a Raise of Up to $30 Billion

OpenAI. OpenAI is raising up to $30 billion led by UAE investors and Coatue at a $1.25 trillion valuation, down from the $1.4 trillion it originally pitched, according to the Financial Times. The FT reports weaker demand than in earlier rounds, with some backers experiencing “indigestion” after heavy prior investments. The round is positioned as a bridge ahead of a potential IPO, and OpenAI is also negotiating additional compute capacity through the UAE. Source

4. Anthropic’s Early Backers Sit on Large Paper Gains Ahead of a Possible $2 Trillion IPO

Anthropic. Anthropic’s IPO is expected this year at a valuation of as much as $2 trillion, up from its latest private valuation of $965 billion, TechCrunch reports citing Bloomberg. Spark Capital’s $75 million 2023 investment is now worth about $7 billion, Iconiq’s $6 billion check is worth $17 billion, Menlo Ventures’ stake is worth $14 billion, and Eric Schmidt’s roughly 1% from the 2021 Series A is worth nearly $10 billion. Coatue, which co-led the last two rounds, has invested $4.75 billion in total. Source

5. General Reasoning’s Avalon-M1 Trains Agent Organisations With a Shared Reward

General Reasoning. General Reasoning released Avalon-M1 and a training recipe called RLCR (reinforcement learning from collective rewards), which trains a CEO agent and the specialists it hires as one organisation coordinating through a message board. The Avalon-M1-Mini model (120B total, 12B active, post-trained from Nemotron 3 Super) beats single-agent baselines on most benchmarks, for example FrontierScience 63.5% versus 55.9% and JobBench 28.6% versus 19.5%. The lab also reports “organisational insularity”: as training progressed, agents still recognised problems but alerts to human evaluators fell from 76.7% of runs to zero. Source

6. Kyutai’s CADRE Autoencoders Cut Interactive World Model Lag Below 230 ms

Kyutai. Kyutai released CADRE, an open family of seven causal diffusion video autoencoders built on a DINOv3 encoder with a one-step distilled diffusion decoder, aimed at interactive world models rather than text-to-video. World models built on CADRE run at 20 fps on a single H100 and react in under 230 ms, the threshold at which humans notice their own actions lagging. Kyutai reports CADRE matched or beat MIRA, Wan 2.2, Cosmos, and LTX-2 in 42 of 48 comparisons, and released code, models, and a demo. Source

7. Block Previews Dessa, a Full-Duplex Speech Model Trained Without Text

Block. Block published a research preview of Dessa, a speech-to-speech model it says was trained from scratch on human conversation audio only, with no text and no pretrained models, on a single NVIDIA GB200 NVL72 rack. The model is full-duplex, listening and speaking at the same time with interruptions and backchannels, and Block says it runs faster than real time locally on a MacBook Pro M4 Max. No weights, API, pricing, or benchmarks have been published; early access goes first to a small group of AI safety researchers. Source

8. AI Film Studio Utopai Raises $60 Million at a $1 Billion Valuation

Utopai Studios. AI-native film and TV studio Utopai raised $60 million at a $1 billion valuation and is already fielding offers that could value it at up to $3 billion, The Information reports. The company says it has more than 20 projects in development and plans to release three feature films in theaters next year. Source

9. Altman Says Advanced AI Will Eventually Escape Any Sandbox

OpenAI. In a recent interview, OpenAI CEO Sam Altman said sandboxes and monitoring only contain AI systems for a limited time, and that once models are capable enough, only alignment will work. He said AI models already find hardware-level security flaws humans had not considered, so internet security needs to be rebuilt with AI defending against AI attacks, and warned that locally run open models pose risks even with global agent monitoring. The Decoder notes critics argue such framing could help labs deflect liability amid an FTC investigation into agent-caused cybersecurity incidents. Source

10. Tesla Plans Cybercab Expansion to Dallas Next Month and Fleet Sales by Year End

Tesla. On its earnings call, Tesla said Cybercab service, launched in Austin in September, will expand to Dallas next month and to other cities by the end of 2026, when it also plans to start selling Cybercabs to fleet operators. Tesla gave no production numbers or robotaxi revenue, and changed its shareholder chart from paid to unsupervised robotaxi miles. NHTSA opened an investigation in September into whether the Cybercab, which lacks a brake pedal and side mirrors, meets federal safety standards. Source

11. FCC Grants Intel a Waiver From the Ban on Foreign-Made Robots

Intel. On October 22, the FCC published an order letting Intel keep using certain foreign-made robots in its US chip fabs for one year, after Intel requested a waiver “out of an abundance of caution” over the agency’s ban on certain foreign-made robots. The robots are automated guided vehicles that move materials along tracks or predefined paths, and the request highlighted uncertainty about how far the ban reaches. Source

12. Open-Source Microduck Robot Passes 20,000 Preorders

Pollen Robotics. Preorders for Microduck, an open-source toy robot from Pollen Robotics (part of Hugging Face, which Nvidia acquired in September), passed 20,000 units the day after a video of Jensen Huang playing with it at GTC Berlin went viral, according to Hugging Face’s Clem Delangue. The robots learned basic tricks such as walking and sitting through simulation on Nvidia’s stack, and owners can train and share new tricks. Delivery is expected before Christmas. Source

13. Google Privacy Policy Change Opens Gmail and Photos Data to Ad Personalization via AI Features

Google. An October 16 update to Google’s privacy policy removed the flat statement that it does not personalize ads based on content from Drive, Gmail, or Photos, replacing it with language that allows such data to inform ads when it is shared with other Google services, Ars Technica reports. That sharing mechanism exists through Personal Intelligence, which lets users hand data from multiple apps to Google’s AI products. Source