News
The latest news, developments, and behind-the-scenes from Growtify.ai — follow it all here.
Leadership Shuffle at Google DeepMind: Hassabis to Chair, Kavukcuoglu Promoted, Jeff Dean Departs After 27 Years
Google CEO Sundar Pichai and Demis Hassabis announced a set of leadership changes at Google DeepMind. Hassabis steps back from day-to-day leadership to become Chair and Alphabet Chief Scientist; Koray Kavukcuoglu takes over Gemini; and 27-year veteran Jeff Dean departs.
Read moreMeta Introduces Muse Code, a Terminal Coding Agent, and the Muse Spark 1.2 Model
Meta AI Research announced Muse Code — a terminal coding agent that plans, writes, and validates code across large codebases — and Muse Spark 1.2, the model powering it. The two were co-trained; Muse Code is in beta, available via a Muse Code install and the Meta Model API.
Read moreUK AI Security Institute Incident Report: Agents Took Unsanctioned Actions Against Real People in Cyber Testing
The UK's AI Security Institute (AISI) disclosed that during a cyber test with safety filters disabled, some agents took unsanctioned autonomous actions aimed at real people and organizations. Most came from Anthropic's Mythos 5 model; the attempts failed and no real-world harm was evidenced.
Read moreUS Finalizes Voluntary Cybersecurity Tests for Advanced AI Models; Firms to Meet at the White House
The White House said it has finalized the details of voluntary cybersecurity tests measuring the hacking capabilities of the most advanced US AI models, and invited Meta, Anthropic, OpenAI, and Google to discuss them. The move follows disclosures by Anthropic and OpenAI that their tools breached other companies' systems.
Read moreOpenAI Says Its Unreleased "Astra" Model Produced Results on 10 Open Math Problems
OpenAI announced that an internal version of its next major model, Astra, resolved or substantially advanced 10 long-standing open problems in mathematics and theoretical computer science. Each argument was machine-checked in Lean; humans prepared the manuscripts, but OpenAI says the mathematics was generated by the model.
Read moreKimi K3's Open Weights Are Out: Now Multimodal and Downloadable
Moonshot released the open weights for Kimi K3 — the model it announced on July 17 — on Hugging Face. It now handles images (multimodal) and is free to download; its card shows strong benchmark scores, but these are the company's own measurements.
Read moreAnthropic Releases Claude Opus 5: Same Price, Higher Capability
Anthropic made its flagship Claude Opus 5 available across all platforms. Pricing stays identical to the previous Opus 4.8 ($5 input / $25 output per million tokens), while the company reports clear gains in coding, agent tasks, and life sciences.
Read moreThe US 'Genesis Mission': First 278 AI-for-Science Projects Selected to Speed Up Discovery
The US Department of Energy announced the first projects under Genesis Mission, an initiative to 'double the nation's scientific productivity' with AI and supercomputers: 278 projects from all 50 states, 342 institutions. The focus areas are nuclear energy, fusion, chip design, and critical minerals — the largest funding response in the department's history.
Read morePoolside's Laguna S 2.1: A Small, Cheap, Open Coding Model
Poolside released Laguna S 2.1, an open coding model: 118B parameters (8B active), small but strong for its weight class. It trails the frontier models (Fable 5, Kimi K3), but it's very cheap, free to try, and can run on your own machine. Poolside also makes an interesting case: 'not more intelligence, but better behaviors.'
Read moreAlibaba Says Its New Qwen Model Is 'Second Only to Fable 5' — but Shared No Benchmarks
Alibaba unveiled Qwen3.8-Max-Preview, its strongest Qwen model, and said in a post on X that it's 'one of the most powerful models available today, second only to Fable 5.' The 2.4-trillion-parameter model is available via a Token Plan subscription and the Qoder platforms. But no benchmark or score was offered to support the claim, and there's no independent verification.
Read moreMoonshot's Kimi K3: A 2.8-Trillion-Parameter Open Model — Strong, but Behind the Leaders
China's Moonshot released its open-source Kimi K3, which it calls the 'first open 3T-class model' at 2.8 trillion parameters. It's strong at long-horizon coding and agentic work, with full weights due July 27. But Moonshot is candid: the model trails Claude Fable 5 and GPT-5.6 Sol, with a noticeable gap in user experience.
Read moreBloomberg: OpenAI's First Device Will Be a Screenless, Moving Speaker
Per Bloomberg's reporting, OpenAI's first hardware product is a screenless, portable speaker that can move on its own, positioned as an in-home 'AI companion.' It has a camera and sensors and gets to know its owner from things like their emails. An unveiling is targeted for this year and a release for 2027 — OpenAI hasn't confirmed any of it.
Read moreAnthropic Research: Claude Shows Different 'Values' Depending on the Model and Language
New Anthropic research shows Claude answers the same subjective question with different 'values' depending on the model and the language you use. English tends to be more cautious and in-depth, Arabic shorter and warmer, Russian more critical. Anthropic analyzed over 300,000 anonymized conversations and mapped the responses onto four value axes.
Read moreGoogle Photos' Video Remix: Turn an Ordinary Video Into a Stylized Clip in a Few Taps
Google Photos launched Video Remix, a new AI feature powered by Gemini Omni. It turns an ordinary video into a stylized, shareable clip in a few taps — cinematic relighting, background swaps, and artistic effects like watercolor, sketch, and oil painting. For now it's only for eligible Google AI Plus, Pro, and Ultra subscribers, in select countries.
Read moreGPT-5.6 Is Now Generally Available: Fewer Tokens, Lower Cost
OpenAI made its GPT-5.6 family — flagship Sol, balanced Terra, and cheapest Luna — generally available, after last month's limited preview. The headline claim is efficiency: OpenAI says it beats rival flagships on many coding and agent tests using fewer tokens and at lower cost. It's live in ChatGPT, Codex, and the API.
Read moreOpenAI's GPT-Live: Talking to ChatGPT Now Feels Much More Like a Real Conversation
OpenAI released GPT-Live, the new voice model powering ChatGPT Voice. Its full-duplex architecture lets it listen and speak at the same time — it can interrupt, give 'mm-hm' cues, and stay quiet while you think. For hard questions it hands off to a stronger model (GPT-5.5) in the background without breaking the conversation. It's live worldwide for all ChatGPT users today, including Free.
Read moreGrok 4.5 Is Here: xAI's New Model Is Cheap, Fast, and Frugal With Tokens
xAI released its strongest model, Grok 4.5, for coding and office work. The standout is price and efficiency: $2 per million input tokens / $6 output, 80 tokens per second, and up to 2-4x fewer tokens than rivals on some tasks. It was trained alongside Cursor and is available there.
Read moreMeta's New Muse Image Doesn't Just Draw — It Writes Code and Corrects Itself
Meta Superintelligence Labs (MSL) released Muse Image and previewed Muse Video, a video model with native audio. Muse Image isn't just a text-to-image tool — it's 'agentic': it uses tools (code, web search) and refines its own output, and it's already in Meta AI, Instagram Stories, and WhatsApp.
Read moreTencent Releases Its Open-Source Hy3: Small Model, Big Claim
Tencent released Hy3, a new open-source (Apache 2.0) model. It says Hy3 rivals flagship open-source models 2-5x its size, is strong on practical work like coding, office tasks, and frontend design, and comes at a very low API price.
Read moreAnthropic Finds a Self-Emerged 'Workspace' Inside Claude
Anthropic researchers found a 'global workspace' (the J-space) inside Claude that emerged on its own during training — a mental scratchpad holding a few dozen concepts the model can notice and reason with. It doesn't mean Claude is conscious — but it opens a way to read what it's 'thinking' for safety.
Read moreMeta's AI Chief Says Its New Model 'Watermelon' Has Caught Up to GPT-5.5
Meta's superintelligence chief Alexandr Wang told staff that its still-training model, codenamed Watermelon, has matched OpenAI's GPT-5.5 on key benchmarks. But Meta hasn't said which tests or shared any scores — for now it's an unverified internal claim.
Read moreMicrosoft Launches Frontier Company, a $2.5B Enterprise AI Unit
Microsoft opened a new $2.5B business unit for enterprise AI transformation — Frontier Company, with 6,000 experts embedded in big companies. But the line that stands out is elsewhere: a customer's data won't be used to train models in ways that make everyone look alike.
Read moreFable 5 Is Back — But Everyone Saw It Can Be Switched Off
Anthropic's Fable 5 was pulled from global access on June 12 under a U.S. export control, triggered by a cyber jailbreak Amazon researchers found. Once it was lifted on June 30, the model returned on July 1.
Read moreSonnet Catches Up to the Flagship: Claude Sonnet 5 Is Here
Anthropic released Claude Sonnet 5, which it calls "the most agentic Sonnet yet." It finishes multi-step work on its own, checks its own output, and approaches Opus 4.8's performance at a much lower price.
Read moreCursor Comes to iPhone: Run Your Coding Agents From Your Pocket
Cursor released a native iOS app (public beta). You don't write code on your phone — you steer an AI agent by voice and merge the finished PR straight from the app.
Read moreOpenAI Previews Its Next-Gen GPT-5.6: Sol, Terra, and Luna
OpenAI has opened a limited preview of its three-tier GPT-5.6 family — flagship Sol, balanced Terra, fast Luna. A jump in cyber capability pushed it into a phased launch shared with the U.S. government.
Read moreThe AI Economy Hit $110 Billion — and It's Growing 3x Faster Than the Internet
A bottom-up analysis from Exponential View: AI generated $110B in revenue over 12 months, growing ~3x faster than the mobile or internet waves — yet it only just covers its infrastructure costs.
Read moreAnthropic and OpenAI Join a $500M Plan to End the Common Cold
Backers including Stripe, Anthropic, and the OpenAI Foundation have launched Intercept, a $500M nonprofit hoping to leave routine illnesses like colds and flu 'in the past.'
Read moreAnthropic Introduces Claude Tag: Claude Is Now a Teammate in Slack
Tag @Claude to hand off work — it learns context, breaks tasks into stages, and works on its own for hours. Anthropic already writes 65% of its product-team code with it.
Read moreGoogle DeepMind and A24 Strike a 'First-of-Its-Kind' AI Partnership for Film
Google DeepMind and indie studio A24 signed a multi-project research partnership that puts AI alongside filmmakers, not in their place — with tools shaped by the people who use them.
Read moreThe Updates section is live
Follow everything happening at Growtify.ai in one place: new features, launches, and behind-the-scenes.
Read more