TOP STORIES
Google DeepMind had its biggest shake-up in years: on 5-6 August, Demis Hassabis stepped back from day-to-day running of DeepMind to become Alphabet chief scientist and DeepMind chairman, CTO Koray Kavukcuoglu took over as CEO reporting to Sundar Pichai, and 27-year Google veteran Jeff Dean left the same week to co-found a research-automation startup, Discovery Loop, with Sanjay Ghemawat, Oriol Vinyals and Quoc Le. This is the clearest signal yet that Google feels behind OpenAI and Anthropic and is restructuring to close the gap. Source: https://www.axios.com/2026/08/05/google-deepmind-demis-hassabis-ai and https://techcrunch.com/2026/08/05/jeff-dean-and-other-top-ai-researchers-are-leaving-google-to-launch-their-own-startup/
Meta agreed to pay up to $17.1 billion to settle a child-safety lawsuit brought by 29 US states over Instagram and Facebook's addictive design for minors, the largest single settlement in the company's history. Meta must also cap under-18 usage, block overnight access, and limit school-hours notifications. It lands the same month Meta raised its 2026 capex guidance to as much as $145 billion, underlining how much regulatory risk is now baked into the AI-era business model. Source: https://www.npr.org/2026/08/26/nx-s1-5944781/meta-settlement-child-safety-lawsuit
Stanford and the Arc Institute published the first complete viral genomes designed entirely by AI, using genome language models Evo 1 and Evo 2 to generate roughly 700,000 candidate bacteriophage genomes, of which 285 were synthesised and 16 proved functional and able to kill drug-resistant E. coli. A companion Science editorial from Johns Hopkins biosecurity researchers warned that governance for this capability "does not yet exist" — a genuinely new category of AI safety concern, not a hypothetical one. Source: https://www.techtimes.com/articles/323516/20260807/ai-authored-functional-viruses-scratch-screening-was-built-living-things.htm
SpaceX completed its $60 billion all-stock acquisition of Cursor-maker Anysphere on 14 August, the largest startup exit on record, folding Cursor into a new "SpaceXAI" division with access to SpaceX's Colossus supercomputer. It is the clearest sign that coding-agent infrastructure is now viewed as strategic compute-and-distribution territory, not just a SaaS category. Source: https://www.bloomberg.com/news/articles/2026-08-14/spacex-completes-its-60-billion-cursor-acquisition
Autodesk closed its largest-ever acquisition, buying maintenance-management platform MaintainX for $3.6 billion on 3 August, then beat Q2 FY27 earnings estimates on 27 August (EPS $3.30 vs $3.12 consensus, revenue $2.05bn) and raised full-year guidance. The MaintainX deal creates a new Autodesk Operations Solutions division pairing Tandem, Flexsim and Fusion Operations with MaintainX's work-order and asset data — Autodesk's clearest move yet to own the "operate" phase of the AEC lifecycle, not just design and construction. Source: https://adsknews.autodesk.com/en/news/autodesk-to-acquire-maintainx-advancing-unified-platform-in-operations/ and https://adsknews.autodesk.com/en/news/5-things-to-know-autodesk-q2fy27-results
FRONTIER MODELS
Google shipped Gemini 3.7 Flash on 13 August, a fast coding-and-agents "workhorse" model with a 1,048,576-token context window scoring 65.3 percent on DeepSWE v1.1, priced from $0.75/M input tokens through year-end — notable because it arrived before the long-delayed Gemini 3.5 Pro. Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/
xAI (now operating as SpaceXAI following the Cursor acquisition) released Grok 4.6 on 12 August, an upgrade to Grok 4.5 at the same 500K-token context class, priced at $2/$6 per million tokens and adding an "xhigh" reasoning tier — Musk has described it as a 1.5-trillion-parameter model with gains coming mainly from better SFT and RL rather than a bigger base model. Source: https://www.unite.ai/spacexai-launches-grok-4-6-for-long-running-agents/
Alibaba moved Qwen3.8-Max to general availability on 3 August with a full benchmark table: a 2.4-trillion-parameter, 95B-active MoE model that Alibaba claims beats or matches Claude Opus 4.8 and GPT-5.6 on several agentic and research-reproduction benchmarks (93.0 vs 90.5 vs 88.8 on PaperBench). Treat the head-to-head framing as vendor-reported until independently reproduced. Source: https://www.bloomberg.com/news/articles/2026-08-03/alibaba-drops-another-china-ai-model-with-breakthrough-performance
DeepSeek took V4-Pro out of preview and into general availability on 12-13 August (DeepSeek-V4-Pro-0813), a 1M-context, agent-focused model, while also introducing peak/off-peak pricing that roughly quadruples V4-Pro output cost at peak hours. Source: https://tech.yahoo.com/ai/articles/deepseek-officially-launches-v4-pro-181255468.html
Z.ai released GLM-5.3 on 14 August without retraining its base model, deriving gains purely from scaled post-training — Z.ai reports it lifts Terminal-Bench 3.0 from 4.6 percent to 28.3 percent — then released open weights on Hugging Face on 28 August after a two-week safety review. Source: https://www.marktechpost.com/2026/08/14/z-ai-ships-glm-5-3-without-retraining-the-base-model-better-at-complex-coding-and-long-horizon-tasks/
AEC, BIM & CONSTRUCTION TECH
Autodesk's largest-ever acquisition — MaintainX for $3.6 billion, closed 3 August — creates a new Autodesk Operations Solutions division and signals Autodesk is pushing hard into the "operate" phase of the building lifecycle, not just design and construction. Full Q2 FY27 results (27 August) beat estimates and came with raised guidance. Source: https://adsknews.autodesk.com/en/news/autodesk-to-acquire-maintainx-advancing-unified-platform-in-operations/
Munich-based reality-capture and spatial-data platform NavVis raised an $85 million Series D in the week ending 17 August, the largest single round in that week's construction-tech funding roundup — a strong signal of continued investor appetite for point-cloud and digital-twin infrastructure. Source: https://bricks-bytes.com/funding-ma/latest-construction-technology-funding-rounds-17th-aug-2026/
buildingSMART International signed a liaison agreement with the Alliance for OpenUSD on 25 August, a step toward bridging IFC-based BIM interoperability with the 3D/USD ecosystem used in visualisation, gaming and digital-twin tooling. Source: https://www.geoweeknews.com/articles/buildingsmart-international-signs-liaison-agreement-with-the-alliance-for-openusd/
Two-thirds of finalists in Bentley Systems' 2026 Year in Infrastructure Awards used AI in their submitted projects, a marked jump from 2025, according to finalist coverage published 20 August — a useful, if self-reported, data point on real-world AI adoption in infrastructure engineering rather than vendor marketing claims. Source: https://highways.today/2026/08/20/bentley-yii-2026-finalists/
AGENTS & DEVELOPER TOOLS
Claude in Chrome went generally available to all paid Claude plans on 26 August, letting Claude take autonomous actions in the browser — clicking, typing, navigating, filling forms — validated by a safety classifier per action rather than requiring approval each time. Source: https://claude.com/blog/claude-in-chrome-generally-available
GitHub Copilot's August releases added Kimi K3 and MAI-Code-1.1-Flash as selectable models, shipped Agent Plugins 1.0 generally in VS Code, Copilot CLI, the SDK and the Copilot app, and enabled organisation-level custom agents that Visual Studio surfaces automatically to members. Source: https://github.blog/changelog/2026-08-13-github-copilot-in-visual-studio-august-update-2/ and https://github.blog/changelog/2026-08-28-github-copilot-weekly-releases-august-24/
Cursor — now under SpaceX ownership following the $60bn acquisition that closed 14 August — shipped Google Workspace plugins giving agents direct Gmail, Drive and Calendar access, and moved Cloud Agents to no longer require a connected GitHub or other SCM provider to start prompting. Source: https://www.bloomberg.com/news/articles/2026-08-14/spacex-completes-its-60-billion-cursor-acquisition
OpenAI's Assistants API is being shut down; Zapier and other automation platforms deprecated all actions built on it, with Zaps that still use it stopping on 26 August unless manually migrated to the Responses API. Anyone with older Zapier-to-ChatGPT automations should check this now rather than discover it broken. Source: https://www.techtimes.com/articles/325345/20260824/openai-assistants-api-shuts-down-tuesday-no-automated-migration-threads-risk.htm
OPEN SOURCE & LOCAL AI
Meta returned to open-weight releases with Muse Glimmer on 10 August: a 30-billion-parameter, Apache 2.0-licensed agentic model distilled from Muse Spark, quantised to roughly 4-bit so it fits under 20GB of VRAM, with a claimed 3.1x generation speed-up on an RTX 5090 via speculative decoding. Source: https://venturebeat.com/technology/meta-returns-to-open-source-with-muse-glimmer-an-apache-2-0-licensed-30b-parameter-ai-model-optimized-for-agents-available-now
MiniMax open-sourced the base weights of its H3 (Hailuo 3.0) omni-modal video model on Hugging Face on 3 August under a Community License that explicitly excludes local deployment in the US, EU, UK and South Korea — a reminder to check licence terms before self-hosting, not just capability. Source: https://www.minimax.io/news/minimax-h3-open-source
Z.ai's GLM-5.3 open weights landed on Hugging Face on 28 August after a two-week safety review, following its 14 August API launch — see Frontier Models above for benchmark detail. Source: https://www.marktechpost.com/2026/08/14/z-ai-ships-glm-5-3-without-retraining-the-base-model-better-at-complex-coding-and-long-horizon-tasks/
IMAGE, VIDEO & AUDIO
Black Forest Labs took FLUX 3 Video to general availability on 4 August via the BFL API: 20-second clips with audio generated alongside the frames rather than dubbed afterward, in text-to-video, image-to-video and video-continuation modes, from $0.06/second in draft mode. Source: https://the-decoder.com/black-forest-labs-makes-flux-3-video-generally-available-and-claims-it-beats-seedance-2-0/
Alibaba launched a beta of HappyShrimp 1.0 on 17 August, an AI music-generation model that produces full vocal or instrumental tracks — melody, arrangement, lyrics and vocals — from a text description of an emotion, story or genre. Source: https://www.bloomberg.com/news/articles/2026-08-17/alibaba-launches-ai-music-generation-model-happyshrimp-in-beta
SOCIAL & COMMUNITY
Jailbreak researcher Pliny the Liberator published an uncensored fine-tune of Alibaba's newly-released Qwen3.8-27B on 20 August; developer Alex Finn then posted that the model answered a synthesis question with no jailbreak prompt required on an ordinary Mac the next day, reigniting the open-weights-safety-tradeoffs debate that follows every major open model release. Source: https://www.explainx.ai/blog/pliny-qwen3-8-27b-obliterated-alex-finn-mac-august-2026
A claim that "researchers find AI agents can infect one another with self-propagating mind viruses" went viral on X starting 19 August; the real source was a legitimate arXiv paper co-authored by Anthropic Fellows researcher Jack Lindsey (see Notable Research below), and the viral framing overstated findings that were largely about benign idea-propagation, not autonomous malicious spread. A good case study in checking the paper before sharing the headline. Source: https://arxiv.org/abs/2608.10218
INDUSTRY & BUSINESS
Anthropic signed a $45 billion, six-year compute-rental deal with Nscale on 26 August for roughly 460 megawatts at Nscale's Monarch Compute Campus in West Virginia — Anthropic's eighth major compute corridor and part of a build-out reported at $71 billion total, with capacity coming online in late 2027. Source: https://techcrunch.com/2026/08/26/anthropic-continues-compute-gobbling-streak-in-45-billion-deal-with-nscale/
SpaceX completed its $60 billion all-stock acquisition of Cursor's parent Anysphere on 14 August — see Top Stories for detail. Source: https://www.bloomberg.com/news/articles/2026-08-14/spacex-completes-its-60-billion-cursor-acquisition
Meta agreed to pay up to $17.1 billion to settle state child-safety litigation on 26 August, even as it raised 2026 capex guidance to as much as $145 billion — see Top Stories. Source: https://www.npr.org/2026/08/26/nx-s1-5944781/meta-settlement-child-safety-lawsuit
POLICY & SAFETY
EU AI Act Article 50 transparency obligations took effect on 2 August, requiring providers and deployers to make clear when people are interacting with AI or AI-generated content; non-compliance can trigger fines up to €15 million or 3 percent of global turnover. Source: https://commission.europa.eu/news-and-media/news/safer-and-more-transparent-ai-2026-08-02_en
Stanford and Arc Institute researchers published the first fully AI-designed, functional viral genomes on 6 August, with a companion Science editorial from Johns Hopkins biosecurity specialists warning that oversight for this capability does not yet exist — see Top Stories for full detail. Source: https://www.techtimes.com/articles/323516/20260807/ai-authored-functional-viruses-scratch-screening-was-built-living-things.htm
NOTABLE RESEARCH
OpenAI's unreleased Astra model generated machine-verified solutions to 10 previously unsolved problems in mathematics and theoretical computer science, publishing a 249-page manuscript plus zero-"sorry" Lean 4 proof certificates on GitHub, at a reported compute cost of about $2,000. The most notable single result is an explicit construction of a non-sofic group, closing a question open since 1999. This is a real, verifiable research output — full formal proofs — not a marketing claim, which is what makes it worth taking seriously. Source: https://www.forbes.com/sites/jonmarkman/2026/08/03/openais-astra-solved-10-decades-old-math-problems-for-just-2000/
Researchers including Anthropic Fellows Program member Jack Lindsey published "Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems" on arXiv on 10 August, showing that ideas seeded into one agent in a multi-agent system can propagate to other agents across both live collaboration and context-wiped chains, with benign payloads spreading more reliably than misaligned ones — and that a simple system-prompt warning conferred near-total immunity. Practical implication for anyone chaining agents: prompt hygiene matters more than most teams currently treat it. Source: https://arxiv.org/abs/2608.10218
WORKFLOW AUTOMATION
OpenAI's Assistants API shut down on 26 August with no automated migration path for several action types; Zapier deprecated all ChatGPT (OpenAI) actions built on it, auto-migrating only "Conversation With Assistant (Legacy)" Zaps to the newer Responses-API-based action and leaving them switched off pending manual review — anyone with older Zapier/Make/n8n flows calling the Assistants API needs to check and rebuild them now. Source: https://www.techtimes.com/articles/325345/20260824/openai-assistants-api-shuts-down-tuesday-no-automated-migration-threads-risk.htm
NEW TOOLS & LAUNCHES
Anthropic shipped Claude in Chrome to general availability on 26 August, its first fully autonomous, action-taking browser agent for all paid plans — see Agents & Developer Tools above for detail. Source: https://claude.com/blog/claude-in-chrome-generally-available