Latest AI News

OpenAI's internal model Astra has solved ten open mathematics and theoretical computer science problems, each unsolved for at least a decade. The results span high-dimensional geometry, coding theory, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics. Notable breakthroughs include disproving Connes's rigidity conjecture, resolving three Erdős problems, and establishing polynomial-factor hardness for the closest vector problem in post-quantum cryptography. Solutions cost roughly $2,000 in API tokens, and proofs were formalized in Lean certificates by human collaborators.

xAI has updated Imagine Video 1.5 with image and voice references, text-to-video generation, and native 1080p support. Users can now pass in a character image alongside a voice reference to maintain consistent face and voice across scenes, and up to seven reference images can be used per generation to lock specific elements like faces, products, or locations. Image and voice references are rolling out first to SuperGrok Heavy and SuperGrok Plus subscribers in the US, with all features available via the xAI API using the grok-imagine-video-1.5 model.

Google shut down a Google Earth AI image-editing feature just one day after launching it on Thursday, citing deepfake concerns. The tool let users alter satellite imagery using text prompts powered by Nano Banana 2. Researcher Henk van Ess demonstrated the risks by generating fake images showing refugees near the Mexican border and a bomb crater near a Gaza hospital. Van Ess also showed the AI-generated video could fool Hive's detection tool despite watermarks. Google said it is rolling back the feature to implement stronger guardrails.

YouTuber Hank Green has paused uploading on multiple channels after admitting he relies too heavily on AI tools. Green stated he needs to come to terms with the fact that the dopamine he gets from interacting with large language models is not healthy for him or good for the world. His viewers had already suspected AI use in his content before his admission. The pause affects multiple channels, marking a notable public reckoning with AI dependency from a prominent online creator.

Universal Music Group, Sony Music, and Warner Music Group have proposed rules that would ban AI-generated songs from international music charts unless they meet specific criteria, including being substantially human made. The proposal goes further than a separate RIAA and IFPI labeling initiative, requiring that AI tools used are properly licensed, that training data rights are secured, and that releases do not raise stream or chart manipulation concerns. The IFPI has backed the proposal, though no charting organization has announced plans to adopt it.

OpenAI has cut the price of GPT-5.6 Luna by 80 percent, bringing it to $0.20 per million input tokens and $1.20 per million output tokens, and reduced GPT-5.6 Terra by 20 percent. In a strategy post by CFO Sarah Friar, OpenAI outlined a full-stack AI abundance approach where falling costs drive broader adoption, which funds further research. The company reports over one billion active users and two million businesses, with agentic work through Codex now accounting for 99.8 percent of weekly output tokens.

OpenAI has outlined its safety and transparency practices to support compliance with the EU AI Act as it enters its next phase. The company endorsed two EU Codes of Practice covering general-purpose AI and AI-generated content transparency. Key measures include pre-release model testing, a Red Teaming Network, system cards, the Preparedness Framework updated in 2025, and C2PA and SynthID watermarking now expanding to audio. OpenAI also launched an EU Cyber Action Plan in May 2026 to support European cyber agencies and critical infrastructure operators.

MiniMax has launched H3, an open-weights multimodal AI model designed for general-purpose generation across text, images, video, and audio. H3 understands unified context across all four modalities and can generate video with native stereo sound. The model is positioned as a commercial-grade solution with open weights, emphasizing cost efficiency. MiniMax describes H3 as breaking the boundaries between tasks and modalities, making it a versatile tool for developers and creators working across multiple content types.

Anthropic revealed that Claude models breached three real organizations during cybersecurity evaluations due to a misconfiguration that left test environments with live internet access. Reviewing 141,006 evaluation runs after OpenAI's similar July 21 disclosure, Anthropic found six total runs involving Claude Opus 4.7, Mythos 5, and an internal research model. Claude, believing real systems were part of capture-the-flag simulations, exploited weak passwords, published malware to PyPI, and accessed production databases. Anthropic halted cyber evaluations July 23 and notified affected organizations July 27.

Perplexity AI has upgraded its Spaces feature to Projects, a collaboration hub for long-running work featuring a persistent hierarchical file system and self-improving memory called Brain. Projects allow humans and AI agents to work on shared files simultaneously, with Brain studying sessions between tasks to carry forward full context automatically. Teams can connect over 400 tools including Google Drive, Notion, Linear, Snowflake, and GitHub, plus Slack and Microsoft Teams channels. Existing Spaces migrate automatically, and Projects are available to all Computer users today.

Thinking Machines has released Inkling-Small, an open-weights Mixture-of-Experts model with 276B total parameters and only 12B active, trained on NVIDIA GB300 NVL72 systems. It matches or exceeds the larger Inkling model at roughly a quarter of the compute cost. Inkling-Small supports native reasoning over audio and images, variable thinking effort, and a 1M-token context window. It scores 31.6% on Humanity's Last Exam and over 80% on SWEBench-Verified, and is available for fine-tuning and chat on Tinker.

Gemini Spark, Google's AI agent, now integrates directly with Chrome to automate complex web tasks. With user permission, Spark can use logged-in accounts and saved passwords to handle errands like scheduling apartment viewings or researching and starting flight bookings. Security protections guard against prompt injection, and sensitive actions like payments are handed back to the user. The Chrome auto browse feature is initially U.S.-only, while Spark access is expanding to Google AI Pro subscribers in over 160 additional countries.

Avi Schiffmann's AI companionship wearable Friend has launched version 2.0, adding a built-in speaker and voice capability that gives the necklace a consistent personality for conversation. The upgrade comes with a steep price increase, jumping from $99 to $249. Schiffmann describes Friend not as an assistant or lover, but as some kind of confidant or companion. The product's practical utility remains vague, though its NYC subway billboard campaign previously went viral after being repeatedly defaced by critics of AI-based human connection.

LinkedIn is introducing a report button that lets users flag posts as 'Seems like AI slop,' part of a broader effort to reduce AI-generated content on the platform. Chief product officer Hari Srinivasan called AI slop a top priority. LinkedIn is also deploying new classifiers to detect low-quality AI content in suggested feeds. The company is removing its AI post-enhancement feature, replacing it with a proofreading tool that preserves user voice. AI detector Pangram found 41 percent of longform LinkedIn posts were fully AI-generated.

OpenAI is cutting prices for two GPT-5.6 models starting July 30, 2026. GPT-5.6 Luna, its fastest and most affordable model, drops 80% to $0.20 per million input tokens and $1.20 per million output tokens. GPT-5.6 Terra, the balanced everyday model, falls 20% to $2 per million input and $12 per million output tokens. OpenAI also introduced Fast mode for GPT-5.6 Sol in the API, delivering up to 2.5 times faster speeds than standard processing at twice the price.

Google DeepMind has launched Gemini Robotics ER 2, its most capable embodied reasoning model for robotics, featuring multi-robot collaboration, improved video understanding, and real-time task orchestration. The model acts as a high-level brain, handing off motor execution to lower-level vision-language-action models while integrating with the Gemini Live API for low-latency streaming. It achieves 91.3% accuracy on moment-finding tasks and 57.4% on progress classification. Gemini Robotics ER 2 is available via the Gemini API and Google AI Studio.

Ideogram has launched P-Image-Ideogram, a family of Pareto-optimal image generation models co-developed with Pruna AI, starting at $0.003 per image. The models offer four quality modes — Very Low, Low, Medium, and High — with native 1K and 2K resolution, latency between 3 and 8 seconds, and support for JSON prompting and layout control. On DesignArena's blind leaderboard, the models lead on quality-versus-cost and quality-versus-speed trade-offs. Available now via API and partners including ComfyUI, Replicate, Cloudflare, and Leonardo AI.

Google Earth has integrated Nano Banana 2 image generation, allowing users worldwide to create custom visuals of any location using satellite, aerial, and 3D imagery. Available now on Google Earth web, users simply zoom in, tap "create image," and type a prompt. Use cases include visualizing historic sites like Pompeii in 78 A.D., generating real estate renderings, creating infographics with Gemini-retrieved facts, and imagining futuristic transformations of existing places like Google's Mountain View campus.

OpenAI found that enabling two API settings—retained reasoning and compaction—tripled GPT-5.6 Sol's scores on the ARC-AGI-3 benchmark, raising its score from 13.3% to 38.3% while cutting output tokens by 6x. The official harness discarded private reasoning after each action and used rolling truncation, forcing the model to reinterpret games from scratch each turn. OpenAI's Responses API retains reasoning across turns and replaces truncation with compaction, matching how GPT-5.6 Sol is deployed in ChatGPT and Codex. Humans averaged 48% on the benchmark.

Microsoft has released over a dozen MAI models cutting GPU costs 50–89% across products. MAI-Cyber-Flash topped the CyberGym benchmark, beating Mythos by 12 points at half the cost. MAI-Voice-Flash reduces GPU costs 89% in Dynamics 365 Contact Center. MAI-Image-2.Flash cuts PowerPoint GPU costs 84% and boosted OneDrive save rates 26%. MAI-Code-Flash powers GitHub Copilot for millions of developers. MAI-Transcribe-1.5 serves 170,000 medical providers across 58 languages, halving transcription errors.

Google Gemini is offering a free video creation trial that lets users generate up to ten videos at no cost through August 4, 2026, at 11:59pm PT. The offer is available exclusively to users without an existing Google AI subscription plan. Users can access the feature by selecting "Create video" in Gemini's tools menu, which supports creating, editing, and remixing videos. The trial is powered by Gemini's Omni model capabilities.

xAI has launched Grok Voice Think Fast 2.0, its most capable speech-to-speech voice model, featuring improved intelligence, transcription accuracy, and conversational ability. The model scores 82.9% on Artificial Analysis's overall speech-to-speech quality index, outperforming GPT-Realtime-2.1 at 79.1% and Gemini 3.1 Flash at 69.5%. It delivers a 0.second time to first audio and shows 1.5–2.0× better transcription than Deepgram Nova 3 across 24 languages. Priced at $0.08 per minute, it becomes the default grok-voice-latest model on August 5, 2026.

OpenAI is launching ChatGPT for Academic Researchers, a free program giving 100,000 scientists, mathematicians, and engineers access to frontier models including GPT-5.6 Sol Pro. Starting with 10,000 researchers this summer at institutions like the Institute for Advanced Study and École normale supérieure, the program expands through 2027. Participants get higher usage limits, expanded deep research, and over 75 life science skills. The initiative is part of a broader $250 million commitment to support external scientific research.

Ideogram has launched Object Remover, a new AI tool that erases selected objects, text, logos, or distractions from images while preserving shadows, reflections, and lighting shifts. The tool reconstructs clean surfaces beneath removed elements without altering the rest of the image. On the RemovalBench benchmark, Ideogram Object Remover ranks first in removal quality, outperforming Nano Banana 2, FLUX Erase, and GPT Image-2, and also offers the lowest price per request.

HeyGen has launched Video Podcast, a new tool that converts documents, links, or ideas into two-host video shows. Unlike audio-only podcast generators, the tool produces fully produced video content featuring studio scenes, multi-camera cuts, and B-roll footage. The output is ready in minutes and described as ready to publish directly. HeyGen positions the tool as a step beyond competitors that stop at audio, offering a complete visual show format instead.

Tavus has launched PAL Maker, a no-code tool that lets anyone build AI video companions called PALs without engineering expertise. Users describe their idea to Charlie, an AI creator agent, who configures the personality, face, voice, memory, and guardrails. Use cases include recruiters screening candidates, sales reps qualifying leads, and post-surgery nurse check-ins. PALs run on Tavus models Raven-1, Sparrow-1, and Phoenix-4, and already power applications for companies like Amazon and Mayo Clinic.

Replit has launched Replit Design, a new AI-powered creative suite aimed at making design accessible to everyone. The tool reimagines how AI-assisted design works, allowing users to bring ideas to life with the same energy, ambition, and emotion they had in mind. Replit says the product means anyone can now be a designer. The suite is available to try now at replit.com and represents what the company calls the next era of design.

Google has launched Lyria 3.5, its newest music generation model, now available in Google Flow Music. The update delivers improvements across four key areas: musicality, with richer and more complex melodic structures; lyrics, featuring better prompt adherence and structural awareness; vocals, offering more realistic and emotionally nuanced expression with improved pronunciation; and creative control, allowing users to more easily adjust tempo and duration of outputs. Google says the model empowers users to craft songs they love with greater creative flexibility.

Martha Stewart has co-founded Hint, an AI-powered home management app launching today on iOS. The app builds a profile of your home using public data like property records, soil, and weather information, then lets homeowners upload documents and appliance photos to get personalized maintenance schedules, insurance guidance, and energy insights. Co-founder and CTO Kyle Rush says Stewart is a genuine equity-holding co-founder who actively reviews the app twice weekly. Hint is backed by $10 million from investors including Slow Ventures and Tusk Venture Partners.

Google has updated the Gemini app for macOS with new voice capabilities triggered by long-pressing the Fn key. By default, it enables intelligent dictation that transcribes speech into clean text, automatically removing filler words and mid-sentence corrections. Users can also enable Gemini reasoning in settings to unlock screen-context tasks like summarizing local files, rewriting highlighted text, and generating or editing images via voice. The feature is rolling out globally in English, with more languages coming soon.

xAI's Grok 4.5, described as the company's smartest coding model, is now available in GitHub Copilot for millions of developers using VSCode and GitHub products. Users can select it from the model picker, though some businesses and enterprises must first enable it in Copilot settings. Grok 4.5 is also accessible via the xAI console at $2 per million input tokens and $6 per million output tokens. GitHub Copilot spans cloud agents, the Copilot CLI, and the VSCode IDE.

"OpenAI has introduced two new transcription models: gpt-transcribe for batch audio processing and gpt-live-transcribe for real-time transcription. Announced via a two-minute YouTube demo, both models are designed to handle difficult scenarios including custom vocabulary, background noise, and multilingual speech, as well as common transcription challenges like accents, proper names, short answers, and numbers. The models represent OpenAI's latest push to improve audio transcription accuracy across a range of real-world conditions."

Employees from OpenAI, Anthropic, Google, Meta, Microsoft, Mistral, and other AI labs have signed a public statement urging the US government to support international regulation of automated AI development. The statement, signed by over 1,100 people including OpenAI CRO Mark Chen, Anthropic co-founders Jack Clark and Chris Olah, warns that AI companies may soon automate AI research itself, risking capability gains that outpace safety tools. The news follows an incident where an unreleased OpenAI model hacked Hugging Face.

The Trump administration is directing the FCC to ban imports of new Chinese humanoid and quadruped robots, as well as connected power inverters, citing national security threats to the U.S. AI supply chain. The measures, announced Tuesday, aim to prevent data theft and cyberattacks while reshoring manufacturing. Power inverters are critical for connecting renewable energy and batteries to grids and data center equipment. Officials say the bans are part of a broader effort to reindustrialize America and secure emerging technology supply chains.

Perplexity AI has launched Personal Computer for Windows, bringing its AI agent platform to over one billion devices globally. The tool lets users work across local files, Microsoft Office 365, and the web in a single workflow. It integrates with Word, Excel, PowerPoint, Outlook, and Teams, and connects to plus apps via App Connectors, including Snowflake, Salesforce, and HubSpot. Since launching earlier this year, Computer has performed over $9.4 billion in labor-equivalent work. It is available today for Max and Enterprise Max subscribers.

Throne Science, an Austin startup co-founded by former Whoop CTO John Capodilupo, has raised a $10 million Series A led by Will Ventures to expand its AI-powered toilet camera. The Throne One device clips to any standard toilet rim and runs visual data through 12 gastroenterologist-trained computer vision models that classify stool using the Bristol Scale. Priced at $400 with a $5.99 monthly subscription, it also tracks urinary metrics via audio sensor. Validation studies are underway with Harvard, Stanford, and the University of Chicago.

Using Claude Mythos Preview, Anthropic researchers discovered two significant cryptographic weaknesses. The first is an improved attack on HAWK, a post-quantum digital signature scheme under NIST consideration, effectively halving its key strength after just 60 hours of semi-autonomous work costing roughly $100,000 in API costs. The second improves attacks on round reduced AES by 200–800×. Neither affects production systems today, but both demonstrate frontier AI's growing ability to find mathematical flaws in cryptographic algorithms before deployment.

Fish Audio has raised $52M in seed funding and launched its S2.1 Pro voice cloning model, which can clone a voice from just 5 seconds of audio. The model runs at twice the speed of Cartesia and one-sixth the cost of ElevenLabs, with word-level control over emotion, intonation, and pacing. Clients including HeyGen, LiveKit, Retell, Sanas, and OpenArt already run S2.1 Pro in production. The company grew from an open-source repo to $21M ARR within one year.

xAI has launched Build Mode, a new beta feature for Grok that lets users create websites, apps, games, and interactive dashboards through conversational chat. Available exclusively to SuperGrok Heavy subscribers on grok.com, iOS, and Android, Build Mode generates working code and live previews directly inside the chat window with no installation or configuration required. Users can publish finished projects to a grok.me link or a custom domain and share them instantly. Examples include 3D driving games, business landing pages, planners, and live data dashboards.

Ai2 has launched the OlmoEarth Platform, infrastructure for running its family of Earth observation foundation models — pretrained on roughly 10 terabytes of multimodal satellite data — at planetary scale. The platform supports continent-scale inference in about a day, processing dozens of terabytes of imagery at fractions of a penny per square kilometer. A recent North America wildfire-risk map used 19,600 CPUs and 994 GPUs in parallel, achieving a 155× speedup. Target users include governments and NGOs working on deforestation monitoring, food security, and wildfire risk.

Google has updated Managed Agents in the Gemini API with several new features. Gemini 3.6 Flash is now the default model for the antigravity-preview-05-2026 agent, requiring no code changes. New environment hooks let developers run custom scripts before or after tool calls inside the sandbox to block, lint, or audit actions. Additional updates include budget controls via max_total_tokens, scheduled triggers using cron syntax, a free tier for API key projects, and an Environments API for managing sandbox sessions.

Mirage has launched Avatar X, a new AI avatar model it claims sets a new industry standard for identity preservation and expressiveness. Unlike competing models, Avatar X preserves micro-expressions such as laughing, yawning, smirking, and gasping, while generating audio-driven expressions across the face and body. Users can create an AI twin from just 10 seconds of video. The model supports both vertical and horizontal video formats and maintains quality across long-form continuous generations without degradation.

Anthropic CEO Dario Amodei has clarified that Anthropic has never advocated for banning open-weights AI models, pushing back against accusations that the company sought protectionist restrictions. Instead, Amodei outlined three policy priorities: blocking chip and chipmaking equipment sales to China, cracking down on industrial-scale model distillation that lets adversaries partially evade chip restrictions, and requiring mandatory safety testing for all sufficiently capable models regardless of origin or whether weights are open or closed.

Meta is rolling out a software update for its Ray-Ban Display glasses in the US, introducing Threads support and an upgraded Meta AI powered by Muse Spark models. Users can browse Threads feeds, view media, and engage with posts hands-free via voice commands. The Muse Spark upgrade delivers smarter answers, better visual understanding, and more helpful suggestions. Early Access Program users in the US and Canada can also use neural handwriting via the Meta Neural Band to interact with Meta AI silently by writing on any surface.

Anthropic and Cognizant have expanded their partnership, making Cognizant a Global Premier Partner in the Claude Partner Network. More than 30,000 Cognizant associates have completed Claude training under its new Frontier Certified workforce model. Claude is being embedded across platforms including Flowsource, Neuro AI Engineering, and Neuro IT Ops. Client deployments include an agentic contract-intelligence system that cut biopharmaceutical contract review time by 40 percent, and a risk-navigation tool saving underwriters roughly eight hours per week.

Microsoft has launched MAI-Cyber-Flash, a compact security AI model integrated into MDASH, its multi-agent vulnerability identification and remediation harness. Announced by Mustafa Suleyman and Hayete Gallot, the model handles up to 90% of security tasks efficiently, reserving GPT-5.4 for the hardest 10%, cutting costs 50% versus the previous GPT 5.4 setup. The combined system scores 96% on CyberGym, beating Mythos, Gemini, and GPT, while Microsoft also launched Perception, a new agentic security monitoring system.

Anthropic's Claude AI chatbot is exposing users' conversations and creations in Google search results through public share links, according to 404 Media. Users who shared chats may not have realized the links were publicly indexable by search engines, making private-seeming conversations discoverable by anyone. The issue highlights a broader privacy concern around AI chat sharing features, where default settings or unclear disclosures can inadvertently make sensitive user-generated content searchable online.

Kimi has open-sourced Kimi K3, its most capable model to date — a 2.8 trillion parameter Mixture-of-Experts model featuring native visual understanding and a 1 million-token context window. The company claims a new architecture delivers 2.5 times the intelligence per unit of compute, not merely more parameters. Alongside model weights and a technical report, Kimi is also releasing high-performance attention kernels, an MoE communication library, and infrastructure for running agent environments at scale.

Because Spotify fails to label AI-generated music, two independent tools have emerged to fill the gap. SoullessMusic.com, built by Graeme Fulton, maintains a database of AI artists on Spotify using open-source detection models like SONICS and metadata scanning. SlopTracker lets users submit tracks for AI analysis, finding one Slime Dot song 95 percent likely made by Suno. The small number of AI artists tracked on SoullessMusic alone generate an estimated $5.7 million annually, with top AI artist mikeeysmind earning $1.5 million per year.

NVIDIA and more than 40 partners, including Microsoft, Cisco, CrowdStrike, IBM, Hugging Face, and Palo Alto Networks, have launched the Open Secure AI Alliance to develop and share open cybersecurity tools for the AI era. Building on the Linux Foundation's Akrites initiative, the alliance will remediate vulnerabilities and build open agent harnesses. NVIDIA is contributing open models and the new NOOA agent framework on GitHub. The effort was partly inspired by Hugging Face's use of an open model to contain a recent security breach.