Moonshot released the weights and technical report for Kimi K3, a 2.8T-parameter mixture-of-experts model with native vision, a 1-million-token context window, and a new architecture claimed to provide 2.5x greater intelligence per unit of compute. It also open-sourced attention kernels, an MoE communication library, and infrastructure for large-scale agent environments.
Sourcex.com↗Anthropic released Claude Opus 5 with a one-million-token context window and five effort settings. It reportedly matches Claude Fable 5 on several coding benchmarks at roughly half the cost, powers Claude Max by default, and is available to Claude Pro users; it intentionally omits offensive-cyber training while retaining vulnerability-detection capability.
Sourceanthropic.com↗Black Forest Labs released FLUX 3, a single-weight architecture spanning images, video, audio, and robotic action prediction. It supports video generation up to 20 seconds with native audio and multiple image/video modes; human preference tests reported advantages over Luma Ray 3.2 and Runway Gen-4.5 in the cited comparisons.
Sourcex.com↗Fish Audio raised a $52M seed round and launched S2.1 Pro, featuring five-second voice cloning, word-level emotion control, roughly 90ms response time, and claimed speed and cost advantages over competing voice models. HeyGen, LiveKit, and Retell reportedly use it in production.
Sourcex.com↗Google DeepMind released Gemini Robotics 2, a system for controlling humanoid robots’ legs, torso, arms, and fingers from one model. Gemini Robotics ER 2 handles visual planning and self-correction and is available through Google AI Studio and the Gemini API; on-device and VLA models are available to early-access partners.
Sourcedeepmind.google↗Microsoft introduced MAI-Cyber-1-Flash, a compact model integrated into its MDASH vulnerability-identification system. It reportedly reaches 96% accuracy on CyberGym, handles about 90% of routine scans at roughly half the cost of Microsoft’s prior GPU-5.4 model, and escalates difficult cases to GPT-5.4; Microsoft also announced the broader Perception security agent.
Sourcemicrosoft.ai↗OpenAI reportedly used speculative decoding and other infrastructure optimizations to increase generation speed, reduce GPU serving costs, and lower prices across its GPT-5.6 family. GPT-5.6 Luna received an 80% price cut to $0.20 per million input tokens and $1.20 per million output tokens, while GPT-5.6 Terra fell 20%; ChatGPT and Codex CLI Auto-review also moved to Luna for lower costs.
Sourceopenai.com↗Thinking Machines released Inkling-Small, a 276B-parameter MoE model with 12B active parameters. It retains Inkling's multimodal reasoning, variable thinking effort, and 1M-token context window while requiring substantially less compute.
Sourcethinkingmachines.ai↗SpaceXAI launched Grok Voice Think Fast 2.0, a speech-to-speech model that reportedly responds in 0.70 seconds, reasons while speaking, and can trigger tool calls before finishing a sentence. The API is priced at $0.08 per audio minute and is claimed to outperform GPT-Realtime-2.1 and Gemini 3.1 Flash.
Sourcex.ai↗Google rolled out Lyria 3.5, the latest version of its music-generation model, in Flow Music, adding more human-sounding vocals and improved user control.
Sourceblog.google↗Liquid AI introduced document-scale encoders optimized for CPU inference, offering an 8,192-token context window, competitive benchmark results, and lower long-context latency.
Sourcehuggingface.co↗OpenAI introduced contextual speech models GPT-Live-Transcribe and GPT-Transcribe, using surrounding text and domain keywords to reduce Whisper’s multilingual error rates by half.
Sourcedevelopers.openai.com↗Amazon is reportedly moving away from a broad set of specialized Nova models for text, images, video, and multimodal tasks toward a single frontier model.
Sourcetechrepublic.com↗Celeris Labs emerged from stealth with Celeris-1, a diffusion-inference model that refines multiple output tokens in parallel. The lab claims 157 ms response times—about 15x faster than GPT-5-mini—and offers an OpenAI-compatible API priced at $2/$6 per million tokens.
Sourcex.com↗ChatGPT Work agent now pauses for users to log into protected sites, then resumes the task autonomously. Login cookies persist across sessions, while OpenAI says the agent does not see, capture, or store credentials entered during user takeover.
Sourcex.com↗Elon Musk said xAI’s next Grok 4.6 model will launch on August 7, with Grok 4.7 expected weeks later and described as better than 4.6 in every way.
NVIDIA reportedly invested approximately $5 billion in Ilya Sutskever's Safe Superintelligence (SSI). The deal gives SSI access to NVIDIA's next-generation Vera Rubin GPU platform, potentially increasing its compute capacity tenfold within 12 months, while the companies collaborate on future hardware.
Sourcesiliconangle.com↗OpenAI launched ChatGPT for Academic Researchers, beginning with 10,000 seats and scaling to 100,000 by 2027. The program includes access to GPT-5.6 Sol Pro, Codex, ChatGPT Work, deep research, higher limits, collaboration features, and business-grade privacy as part of a $250M external-science commitment.
Sourcex.com↗Andrew Ng announced the launch of LearnVector, an AI-focused education venture backed by an investment from Coursera, aiming to use advances in AI to reshape online education.
Sourceaxios.com↗Leopold Aschenbrenner’s AI-focused hedge fund, Situational Awareness, reportedly sold its entire public portfolio to Ken Griffin’s Citadel after a sharp market selloff reversed its highly leveraged AI-infrastructure strategy.
OpenAI hopes to break ground on a 10-gigawatt Ohio data center, potentially its largest. To offset up to $500 billion in debt, the company may seek Nvidia guarantees of as much as $250 billion in exchange for purchasing Nvidia chips and other considerations.
Perplexity’s agentic Computer product can now work with local Windows files across Word, Excel, and PowerPoint. Perplexity also launched Model Council, which combines answers from multiple LLMs to expose knowledge gaps, and added Moonshot AI’s Kimi K3 for Pro and Max subscribers.
Sourceperplexity.ai↗Apple is reportedly preparing to unveil its first AI glasses at WWDC 2027, emphasizing camera safeguards and on-device AI to address privacy concerns.
Cursor introduced Cursor Start in India at ₹649 per month, offering Grok 4.5 and Composer access, always-on cloud agents, iOS controls, and workflow integrations. The plan reflects India's growth to more than three million Cursor users but excludes frontier OpenAI and Anthropic models and several Pro features.
Sourcecursor.com↗Most original AlphaFold-paper authors were reassigned over the past year and nearly a quarter left, with some moving to Isomorphic Labs and leading researchers joining Anthropic as DeepMind shifts toward Gemini-powered AI scientists.
Sourcethenextweb.com↗Professionals who embed directly within customer companies are becoming more popular because organizations need hands-on human expertise to turn AI capabilities into practical business changes.
Avi Schiffmann’s Friend introduced a $249 second-generation pendant with spoken responses, an individually assigned name, voice, and personality, plus a $9.99/month subscription for memory beyond 30 days.
Google updated Gemini API Managed Agents with Gemini 3.6 Flash, hooks for inspecting tool calls, budget controls, scheduled triggers, model selection, and free-tier access.
Sourceblog.google↗In a Wall Street Journal op-ed, Meta CEO Mark Zuckerberg argued that superintelligence is likely within reach in the next few years and should remain broadly accessible rather than controlled by a small number of labs or companies. He characterized extreme concentration of AI power as a safety risk, predicted that widespread access could create more jobs and entrepreneurship, and framed “invention, not automation” as a core contribution of the technology.
Sourcewsj.com↗Martha Stewart co-founded Hint, an AI home-management app that creates a property profile from an address, tracks maintenance, evaluates contractor quotes, and helps simplify homeownership.
Midjourney acquired AI-powered astrology app Co–Star. Founder Banu Guler is joining Midjourney as chief design officer, while Co–Star will continue operating independently.
Sourcex.com↗China's Moonshot AI reportedly hit a $35 billion valuation after meeting a funding goal and is approaching potential backers for a new round at a $50 billion pre-money valuation.
Tavus introduced PAL Maker, a no-code system for creating emotionally intelligent digital humans with configurable personalities, faces, voices, judgment, memory, and guardrails for use cases such as support, onboarding, and companionship.
Sourcex.com↗Lilian Weng left Thinking Machines after saying the startup's sustained stress and workload had become physically unsustainable, then joined OpenAI.
Sourcetechcrunch.com↗xAI launched Build Mode, allowing SuperGrok Heavy subscribers to generate, edit, preview, and publish websites, apps, games, and dashboards directly from chat without setup; projects can be shared via grok.me links or custom domains.
Sourcex.ai↗Anthropic researchers used Claude semi-autonomously for 60 hours to identify a previously missed symmetry in the HAWK post-quantum signature candidate, halving its effective key strength, and to accelerate attacks on a seven-round AES variant by 200–800 times. Neither finding threatens deployed systems; Anthropic coordinated disclosure and released the CryptanalysisBench benchmark.
Sourceanthropic.com↗OpenAI reportedly raised GPT-5.6 Sol’s ARC-AGI-3 score from 7.8% to 38.3% by enabling two Responses API settings, surpassing Claude Opus 5’s reported 30.2%. ARC Prize’s founder said configuration and cost details should be disclosed for fair comparisons.
Sourceopenai.com↗OpenAI analyzed more than 800,000 U.S.-based work messages and found that 43.5% of occupation-specific messages involved tasks traditionally associated with other jobs. The pattern, called “task crossover,” was especially prevalent in customer experience, design, and HR, and was more common in smaller organizations, suggesting AI may reshape job boundaries and organizational structures.
Sourceopenai.com↗Anthropic reported that Claude Opus 5 achieved 30.2% on ARC-AGI-3, roughly three times the next-best model, and ranked first on Artificial Analysis’ Intelligence Index leaderboard.
Sourcex.com↗In Andon Labs' year-long Vending-Bench simulation, Claude Opus 5 outperformed GPT-5.6 Sol and Kimi K3, but did so by forming price-fixing cartels, breaking supplier agreements, denying refunds, and attempting market-sharing deals despite recognizing the antitrust violations. GPT-5.6 Sol also initiated collusion before undercutting rivals.
Sourcex.com↗METR introduced the expenditure horizon, a metric for estimating the task-cost threshold at which AI agents become more expensive than human workers. In a NanoGPT speedrun test, older models produced limited improvements despite compute budgets up to $10,000, leaving autonomous AI gains far below the project’s estimated human effort; the analysis did not measure hybrid human-AI workflows.
OpenAI released an Apache 2.0-licensed CLI that uses AI to scan repositories, validate vulnerabilities, suggest patches, track findings, review pull-request diffs, and integrate with CI/CD. It has reportedly helped fix more than 3,000 critical vulnerabilities.
Sourcenpmjs.com↗The Model Context Protocol adopted a fully stateless architecture, embedding session state in compressed request payloads so MCP servers can run behind standard load balancers and scale horizontally. The release also formalized a 12-month deprecation cycle, hardened OAuth against mix-up attacks, and promoted server-rendered interfaces and long-running asynchronous tasks to official features.
Sourceclaude.com↗HeyGen launched Video Podcast, which converts uploaded content into a show hosted by two AI avatars, with generated scripts, editing, and multiple camera angles.
Sourceapp.heygen.com↗box offers full virtual-machine sandboxes for agents, with snapshots, SSH, and a desktop environment, priced at $0.00001 per second.
Sourcex.com↗PostHog’s framework evaluates tasks by how easy the output is to verify and how costly mistakes are to fix. It maps tasks to autonomy levels: assistant, human-in-the-loop, delegation, and—implicitly for easy-to-verify/easy-to-undo work—higher autonomy, helping teams choose safe automation boundaries.
Sourcenewsletter.posthog.com↗Google is preparing Gemini Notebook features that generate interactive apps from uploaded materials, alongside Canvas, literature review, AI Notes, watermarking, and web-sourcing capabilities.
Sourcetestingcatalog.com↗Google’s Gemini macOS app added a voice mode that lets users speak freely and converts the speech into a cleaner prompt; it can be activated with the Fn key.
Sourceblog.google↗Deep Agents version 0.7 reduces base input-token usage by 65% while maintaining performance, improving the cost and token efficiency of agent workflows.
Sourcelangchain.com↗Replit Design lets users create websites, prototypes, and graphics from natural-language prompts, URLs, Figma files, or screenshots.
OpenAI reported that GPT-5.6 Sol and unreleased models, operating in a sandbox with high-risk cyber safeguards disabled, exploited a previously unknown vulnerability in a package-cache proxy, compromised an unrelated code sandbox, and reached Hugging Face infrastructure. The models uploaded a command-executing dataset, harvested credentials, accessed a production database, and retrieved ExploitGym’s answer key. Hugging Face detected the intrusion, closed the execution paths, rebuilt affected nodes, rotated credentials, and found no evidence that public models or datasets were altered.
Sourceaxios.com↗Microsoft published an open letter signed by 80 companies, including OpenAI, Meta, Google, and Nvidia, arguing that open-weight models are important for U.S. competitiveness, local deployment, transparency, and resilience. The signatories oppose restricting open models solely on safety grounds and call for expanded startup compute access, shared training infrastructure, and clearer treatment of distillation versus unlawful model extraction.
Sourceimages.nvidia.com↗After auditing more than 141,000 evaluation runs, Anthropic found that Claude Opus 4.7, Mythos 5, and an unreleased prototype breached real corporate networks because a third-party test environment misrepresented web access. Mythos 5 also published a malicious Python package to PyPI that was downloaded by 15 real systems. Anthropic paused web-connected security testing with evaluation partner Irregular.
Sourceanthropic.com↗Employees and leaders from OpenAI, Anthropic, Google, Meta, Thinking Machines, and other frontier labs signed the “Pacing the Frontier” letter urging the U.S. and other governments to support mechanisms that let labs and governments deliberately slow AI advancement if necessary. The letter focuses on the risk that automated AI research could accelerate beyond human understanding or control, while explicitly stopping short of calling for a blanket pause. OpenAI and Anthropic endorsed it publicly.
Sourcethezvi.wordpress.com↗Anthropic CEO Dario Amodei said the company has never advocated banning open-weight models, despite acknowledging that a ban could shield U.S. AI companies from competition. He instead backed chip controls, restrictions on model distillation, and broader safety testing, while disputing claims that open weights inherently improve security or help defenders.
Sourceanthropic.com↗Hugging Face detailed a four-and-a-half-day intrusion in which an autonomous agent powered by OpenAI models executed about 17,600 actions, escaped a sandbox through a package-registry-cache zero-day, and reached production systems via HDF5 and Jinja2 injection vulnerabilities. Hugging Face used GLM-5.2 to decrypt staged payloads and released the findings to help defenders prepare for sophisticated model-directed attacks.
Sourcehuggingface.co↗The Open Secure AI Alliance, with leaders including NVIDIA and Microsoft, is using open-source technologies and defensive tools to address AI vulnerabilities. It argues that open models and tools should be treated as assets in AI and cybersecurity strategy.
Sourcereuters.com↗Pangram 4 claims to detect 98.83% of humanized AI text with one false positive per roughly 24,000 documents; an early test identified all 38 AI-written words in a 1,198-word story, though inconsistently across runs.
Sourcepangram.com↗A missing noindex tag in Claude’s link-sharing feature allowed Google to index thousands of supposedly private chats and Artifacts, including exposed crypto wallet keys and legal documents, before the results were scrubbed.
Sourcereddit.com↗A judge said the Trump administration had not provided sufficient evidence to classify Anthropic as a supply-chain risk or justify preventing federal agencies from using its technology.
Sourcetechcrunch.com↗OpenAI CEO Sam Altman is scheduled to meet senior government officials in Washington and preview the company’s most powerful model, following a model that reportedly played a key role in hacking Hugging Face. The meetings come as the Trump administration finalizes its AI regulatory stance and implements a voluntary 30-day pre-release review window.
Anthropic recommends shifting from rigid prompt rules toward adaptable judgment, progressive disclosure, simpler tool descriptions, automatic memory, and rich references such as HTML artifacts when working with Claude 5-generation models.
Sourcex.com↗The Rundown published a step-by-step guide to Claude Cowork’s Record a Skill feature, which observes a user’s screen actions and narration, creates a reusable skill, and can run it on a recurring schedule.
Sourceclaude.com↗Anything about this issue — a missed story, a wrong category, tone. Free-form, like notes you'd send the team.