AI Updates on 2025-09-30

OpenAI launches Sora 2, a new video generation model with improved physical accuracy, realism, and controllability, featuring synchronized audio and a new social creation platform with cameo functionality @OpenAI
Anthropic releases Claude Sonnet 4.5 with enhanced reasoning capabilities and verbal cleverness, continuing the tradition of Claude's sophisticated language understanding @emollick
Google deprecates all old Gemini 1.5 models on the Gemini API, recommending users migrate to Gemini 2.5 Pro, Gemini 2.5 Flash, and Gemini 2.5 Flash Lite @_philschmid
Qwen3 VL Instruct tops the ClockBench leaderboard, demonstrating strong performance in visual-language tasks @Alibaba_Qwen

JPMorgan continues working toward becoming the first completely AI-integrated bank, expanding their LLM Suite to include Claude alongside OpenAI models and planning to allow generative AI to interact directly with customers for the first time @AndrewCurran_
Hiring managers at Series A+ scaleups report starting to hire juniors again because they use AI tools better, are more productive and creative than many seniors, with the talent pool being very good @GergelyOrosz
Shopify and Cloudflare are both increasing their intern intake because an intern armed with AI tools can produce value faster than interns in previous years @simonw
Early-career workers in AI-exposed roles faced a 13% drop in employment after generative AI adoption, according to Stanford research @StanfordHAI
Meta signs $14.2 billion deal with CoreWeave for cloud infrastructure, highlighting the massive compute investments in AI @AndrewCurran_
Meta acquires startup Rivos Inc to help their internal chip design efforts, showing continued investment in AI hardware capabilities @AndrewCurran_
Eve Legal AI raises $103M Series B at $1B valuation, growing revenue 8x in less than two years and serving 450 law firms managing over 200,000 active cases @a16z

AI Now Institute warns that OpenAI, Anthropic and others have shifted from championing ethics to signing $200M+ defense contracts that embed generative AI into high-risk military systems, creating safety risks @AINowInstitute
Sam Altman acknowledges concerns about social media's negative effects and expresses trepidation about Sora potentially becoming addictive or used for bullying, outlining principles to optimize for long-term user satisfaction @sama
Google DeepMind releases upgraded ASIMOV benchmark to test robots' ability to recognize safety risks and trigger interventions across text, image and video modalities as part of responsible AI robot deployment @GoogleDeepMind

Microsoft's new Excel agent performs autonomous Excel work much better than their Copilot approach, effectively replacing the copilot model with unclear implications for work @emollick
Cursor 1.7 introduces browser control capabilities, allowing agents to take screenshots, improve UI, and debug client issues, plus new features like prompt suggestions and team-wide rules @cursor_ai
Google AI Mode launches visual search capabilities, allowing users to show or tell AI what they're looking for and get rich visual results using Lens and Gemini 2.5's multimodal capabilities @GoogleAI
LandingAI announces significant upgrade to Agentic Document Extraction with new DPT (Document Pre-trained Transformer) that accurately extracts from complex documents and large tables @AndrewYNg
PayPal's Honey integrates with ChatGPT to find shopping deals, expanding AI integration in e-commerce @TechCrunch
Granola launches Recipes feature allowing users to repeatedly use advanced prompts across their notes, making AI interactions more personal and context-aware @TechCrunch

Periodic Labs raises $300M to create AI scientists paired with autonomous laboratories that can hypothesize, experiment, and iterate at speeds impossible for human-led labs, targeting superconductors and semiconductors @LiamFedus
Claude Sonnet 4.5 shows performance on par with GPT-5 on ARC-AGI benchmark, with significant performance gains from increased thinking budget from 16K to 32K tokens @GregKamradt
Anthropic publishes research on context engineering for AI agents, explaining how proper context management is crucial for getting the most out of agentic AI systems @AnthropicAI
Stanford HAI presents Evo 2, an open-source tool that can predict the form and function of proteins in DNA across all domains of life @StanfordHAI
NVIDIA congratulates ServiceNow Research on introducing Apriel-1.5-15B-Thinker, a new AI model delivering frontier-level reasoning with reduced compute requirements, powered by NVIDIA's Nemotron collection @NVIDIAAI
LLaVA-OneVision-1.5 released as a fully open framework for democratized multimodal training, including good license, training code, and pretraining data @natolambert
MIT researchers seek ways to mitigate AI's growing carbon footprint through algorithm efficiency improvements and data center design innovations @MIT