Latest AI News

ElevenLabs has launched Avatars, a new feature in ElevenCreative that lets users generate lip-synced talking-head videos from a script and a chosen voice in a single interface. Avatars are persistent visual identities built from reference images or text prompts, reusable across multiple videos with consistent appearance. A new Avatar node in Flows enables automated pipelines for batch-producing ad variants, course content, or localized explainer videos. Avatars is available now on all paid ElevenLabs plans.

Perplexity has integrated Deep Research into its Computer product, allowing users to go from complex research queries to finished deliverables like PDFs, dashboards, and decks in a single workflow. The update introduces Search as Code, an architecture where the model writes programs that run thousands of retrieval steps in parallel. Deep Research now pulls from internal files via App Connectors and premium sources like Statista and PitchBook. Benchmarks including Humanity's Last Exam show improved factual accuracy, citation quality, and analytical depth.

Deezer has launched a free AI music detector tool that scans playlists on competing streaming services including Spotify, Apple Music, SoundCloud, and YouTube Music across 20 platforms total. Users visit ai-music-detector.deezer.com, grant access to their streaming account, and Deezer imports and scans playlists for AI-generated tracks. CEO Alexis Lanternier said the move came because no other platform licensed Deezer's detection technology. Apple and Spotify have instead adopted voluntary artist tagging systems.

Meta previewed new features for its Edits video-editing app at an invite-only creator event in Los Angeles, including an AI assistant and a desktop version. The AI assistant uses creators' Instagram data, such as views and video-retention insights, to suggest content ideas and trending audio. A desktop version will enable more precise editing and seamless mobile sync, directly rivaling CapCut's existing desktop app. Also launching today are a Beta tab for experimental features and expanded audience demographic and engagement-time metrics.

Coinbase has launched Coinbase for Agents, connecting AI agents directly to users' Coinbase accounts to trade, pay, and execute financial workflows autonomously. Available now as an MCP for web-based tools like ChatGPT and Claude, and a CLI for terminal environments like Claude Code, users can set limits on trade size and spending. Crypto spot and derivatives trading are fully enabled, with stocks and prediction markets coming soon. A Coinbase Advisor, an SEC and CFTC registered AI financial advisor, is also built into the Coinbase app.

Anthropic reversed a controversial policy for its Claude Fable 5 model after backlash from the developer community. At launch, Fable 5 silently rerouted questions about cybersecurity, biology, and chemistry to less capable models, and degraded performance for AI development tasks without notifying users. Critics accused Anthropic of using safety measures to suppress competition. The company now says flagged requests will visibly fall back to Opus 4.8, with API refusals including a stated reason. Anthropic said it made the wrong tradeoff and apologized.

OpenAI is acquiring Ona, a cloud execution and orchestration company, to expand its Codex platform with secure, persistent, customer-controlled cloud environments. Codex now serves more than 5 million weekly users, up 400% from earlier this year, and has grown beyond software development into broader knowledge work. Ona has helped 2 million developers work in reproducible cloud environments. The deal will let Codex agents continue running inside an organization's own cloud after sessions end, meeting enterprise security, governance, and credential-scoping requirements.

OpenAI has announced support for the European Commission's Code of Practice on Transparency of AI-Generated Content, a key step in implementing the EU AI Act. Building on work since 2024, when it began adding C2PA metadata to DALL-E 3, OpenAI now applies both C2PA credentials and SynthID watermarks to images from ChatGPT and its API, and offers a public verification tool at openai.com/verify. OpenAI was also the first US company to sign the EU's General-Purpose AI Code of Practice in 2025.

Anthropic has launched Claude Corps, a national fellowship program backed by an initial $150 million commitment, aimed at placing 1,000 early-career fellows at nonprofits across the United States. Fellows will receive a $85,000 salary, benefits, and five hours of weekly Claude training while spending 12 months full-time at host organizations. Partners CodePath and Social Finance will manage employment and program evaluation. At least 400 nonprofits will participate in the first year, including food banks, veteran support groups, and marine conservation organizations.

AI video generation company Runway is expanding its existing partnership with Hollywood studio Lionsgate through a slate of new initiatives. The most significant addition is a joint development program focused on creating original intellectual property together, pushing the collaboration beyond tool licensing into co-creating new content. This move signals a deeper integration between AI technology and major studio operations, representing one of the more ambitious AI-Hollywood creative partnerships announced to date.

Anthropic's Anthropic Institute reports that AI is rapidly accelerating its own development, moving toward recursive self-improvement. As of May 2026, Claude authors over 80% of code merged into Anthropic's codebase, and engineers ship 8x more code per quarter than in 2024. Claude Mythos Preview can sustain autonomous work for plus hours. On open-ended tasks, Claude's success rate hit 76% in May 2026, up 50 points in six months. Anthropic warns full recursive self-improvement could arrive sooner than institutions are prepared for.

Anthropic CEO Dario Amodei has published a major policy essay arguing that AI regulation must urgently catch up with the technology's rapid progress. Citing Claude Mythos Preview's disruption of global cybersecurity as proof that frontier models now carry strategic national consequences, Amodei calls for mandatory third-party safety testing of large models, government power to block unsafe deployments, and strong security standards for model weights. Anthropic is simultaneously releasing a legislative proposal on frontier model testing and a job displacement policy framework.

Google has introduced DiffusionGemma, an experimental open model that generates entire token blocks of text simultaneously rather than token-by-token, delivering up to 4x faster inference on dedicated GPUs. The 26B Mixture of Experts model activates only 3.8B parameters during inference, fitting within 18GB VRAM when quantized. It achieves over 1,000 tokens per second on an NVIDIA H100. Released under Apache 2.0, it targets speed-critical local workflows like inline editing and code infilling, though output quality trails standard Gemma 4.

A new drug called daraxonrasib, developed by Revolution Medicines, nearly doubled survival time for advanced pancreatic cancer patients compared to chemotherapy, with patients living 13.2 months versus 6.6 months and reporting less pain. The drug targets the RAS protein family, long considered undruggable due to smooth, featureless surfaces. AI is now accelerating similar efforts against other undruggable proteins like p53 and MYC, which drive roughly 70 percent of cancers, with researchers using AI to screen millions of compounds for new treatments.

UC Berkeley's Center for Responsible, Decentralized Intelligence has launched Agents' Last Exam (ALE), a new benchmark testing AI on real professional workflows across 55 industries. OpenAI's GPT-5.5 via the Codex harness topped the leaderboard with a 24.0% pass rate, beating Anthropic's newly released Claude Fable 5, which scored 22.0%. The benchmark covers 1,490 tasks including 3D modeling and neuroimaging analysis, uses deterministic grading, and keeps 90% of tasks private to prevent contamination. Even top models score 0.0% on the hardest tier.

NVIDIA Blackwell GPUs with Confidential Computing are now powering confidential inference in Apple's Private Cloud Compute, which is expanding beyond Apple's data centers to Google Cloud. Announced at WWDC, NVIDIA is collaborating with Apple and Google to support next-generation Apple Intelligence features using Apple Foundation Models, including models leveraging Gemini family technologies. NVIDIA's Confidential Computing provides hardware-rooted trust, encrypted communication, and remote attestation, ensuring no one, including system builders, can access users' data or conversations during processing.

Anthropic has launched two public beta features for Claude Managed Agents on the Claude Platform. Scheduled deployments let agents run on a cron schedule, automatically starting new sessions for recurring tasks like nightly data syncs or weekly compliance scans, with no scheduler to build or host. Environment variable vaults securely store API keys for CLI tools, attaching credentials at the network boundary so agents never see raw keys. Supported CLIs include Browserbase, KERNEL, Notion, Ramp, and Sentry. Rakuten, Actively AI, and Ando are already using scheduled deployments.

Google has launched Gemini 3.5 Live Translate, a new audio model delivering near real-time speech-to-speech translation across 70+ languages. Unlike turn-by-turn systems, it generates speech continuously, staying just seconds behind the speaker while preserving intonation, pacing, and pitch. The model is rolling out in Google AI Studio for developers, Google Meet for enterprise users in private preview, and the Google Translate app on Android and iOS. All generated audio is watermarked with SynthID to help prevent misinformation.

Electrify, a media investment company, has partnered with The Rundown, the world's largest AI news newsletter founded by Rowan Cheung. The Rundown has grown to over 2 million subscribers across AI, technology, and robotics coverage. Rowan and his team will retain editorial control while Electrify provides operational support, resources, and expertise to expand into new formats, products, and audiences. Electrify Co-Founder Ian Shepherd called it a trusted knowledge brand built around expertise and credibility.

Midjourney founder David Holz has begun sending out invites for the company's first hardware product launch. Holz noted a limited number of spots remain and encouraged anyone who believes they should have received an invite but hasn't to reach out directly. No details about the hardware itself were disclosed in the announcement. This marks a significant expansion for Midjourney beyond its AI image generation software into physical hardware products.

OpenAI has submitted a confidential draft S-1 registration statement to the SEC, a step toward a potential IPO. The company preemptively announced the filing, expecting it to leak. OpenAI stated it has not decided on timing and noted a public offering may still be a while away, as some planned actions are easier to execute as a private company. The filing preserves the option to go public sooner if that proves best, calling it a complicated set of tradeoffs.

Anthropic is releasing a Swift package that lets Apple developers use Apple's Foundation Models framework to call Claude for complex tasks. The integration works on iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27. Developers can use Apple's on-device models for fast local tasks like summarization, then hand off to Claude for multi-step reasoning, code generation, or web search. The package handles streaming, tool calls, and structured responses, with typed Swift outputs from @Generable annotations feeding cleanly into Claude API calls.

OpenAI co-founders Sam Altman and Jakub Pachocki have outlined a three-part plan for the company's third phase: building an automated AI researcher by March 2028, accelerating broad economic growth, and giving every person on Earth access to a personal AGI. The plan emphasizes distributed power over concentration, calling for international coordination on AI safety. OpenAI also warns that fully automating human decision-making is both dangerous and undesirable, stressing that human judgment must remain central.

Anthropic has launched two new features for Claude connector developers. Published connectors in the directory now have a performance dashboard showing active users, total tool calls, directory rank, health scores, error rates, and latency breakdowns per tool. Usage can also be compared across Claude, Claude Code, and Cowork. Additionally, developers can now submit MCP servers to the directory directly in-app. The observability dashboard is in public beta and requires a Team or Enterprise account with Admin or Owner access.

OpenAI has launched the Economic Research Exchange, a program connecting external researchers with OpenAI's tools and datasets to study AI's economic effects on workers, firms, and institutions. Selected researchers will conduct structured, project-based collaborations with defined milestones and data governance safeguards. OpenAI seeks proposals in fields like labor economics, productivity, and inequality. Applications are open now and close July 5, 2026, with selected researchers notified by July 31. The initiative builds on OpenAI's existing Signals measurement efforts.

At WWDC26, Apple unveiled the next generation of Apple Intelligence and Siri AI, an entirely new version of Siri described as profoundly more intelligent and capable. Siri AI is deeply integrated across iPhone, iPad, Mac, Apple Watch, and Apple Vision Pro, with screen awareness, personal context understanding, web search, and cross-app actions. A dedicated Siri app syncs conversation history via iCloud. Updates also include iOS 27, macOS 27, new parental controls, Safari Notify Me, and photorealistic Image Playground.

Amazon has added an AI-powered custom merchandise design feature to Alexa for Shopping, available now to all U.S. customers via the Amazon Shopping app or Amazon.com. Users describe an idea in a text prompt, and the tool generates a design in seconds for products including T-shirts, hoodies, tumblers, and water bottles. Designs can be shared with friends and family, who can each order their own. Production is handled through Amazon's Merch on Demand service with Prime-eligible shipping.

The American Federation of Musicians has sued Universal Music Group and Warner Music Group in the US District Court for the Southern District of New York, alleging the labels licensed member recordings to AI music generators Suno and Udio without compensating or crediting session musicians. The AFM argues the deals triggered a "new use" provision in its collective bargaining agreement. The union seeks monetary damages and disclosure of which recordings were used for AI training. Both labels denied wrongdoing, calling the lawsuit unproductive amid ongoing negotiations.

Google has upgraded NotebookLM with agentic chat capabilities, code execution, and expanded output formats, powered by Gemini 2.5. Each notebook now includes a secure cloud computer with over 100 curated software skills for deeper analysis. New output formats include PDFs, Excel, PowerPoint, CSV, JSON, and data visualizations. The upgraded system achieved a 65% win rate over the prior version, with a 78.2% win rate in advanced web research. Updates are rolling out to Google AI Ultra and Workspace AI customers globally.

Google announced at WWDC that Apple developers can now call cloud-hosted Gemini models directly through Apple's native Foundation Models framework, available starting with iOS 27, macOS 27, and related platforms. The integration uses Firebase AI Logic via the Firebase Apple SDK, letting developers swap between on-device Apple models and cloud-hosted Gemini with minimal code changes. Gemini is also now integrated into Xcode, offering agentic coding assistance for reviewing code, fixing bugs, and building features without switching tools.

OpenAI CEO Sam Altman and Chief Scientist Jakub Pachocki have outlined a three-part plan to ensure AGI benefits all of humanity. Their goals include building an automated AI researcher by March 2028, accelerating broad economic growth, and giving every person on Earth access to a personal AGI. The post warns against power concentration, calls for international coordination on AI safety, and argues that human judgment remains essential even as AI systems become more capable.

President Trump said he's discussing deals with AI companies where the American people can benefit from AI's success. CNBC reports the Trump administration has been actively discussing taking an equity stake in OpenAI, with some of that equity potentially seeding a Public Wealth Fund proposed by OpenAI itself. CEO Sam Altman has reportedly been discussing a government stake in major AI companies since early 2025. The idea mirrors Trump's prior move taking a 10% stake in chipmaker Intel last year.

Anthropic is partnering with synthetic, computational, and analytical chemists to improve Claude's chemistry capabilities. In a new white paper, Anthropic chemist David Kamber tested Claude Opus 4.7, Opus 4.6, and Sonnet 4.6 against dedicated NMR software ChemDraw and MestReNova across 20 novel compounds. Opus 4.7 matched or outperformed both tools on hydrogen shift prediction, averaging ±0.079 ppm error, and tied MestReNova on carbon prediction. Claude also successfully performed inverse structure elucidation from NMR spectra alone, a task existing software leaves entirely to chemists.

OpenAI appears to have introduced a feature allowing ChatGPT users to send emails directly from within writing blocks in the ChatGPT interface, according to a post from the official ChatGPT account on X. However, the original post text was not available for review, so specific details about supported platforms, setup requirements, or limitations cannot be confirmed. The title suggests this is a notable workflow integration, but a fully accurate summary requires access to the complete source material.

New York's state legislature passed a one-year moratorium on new large data centers, potentially the first statewide ban of its kind in the US. The bill targets facilities with peak demand of at least 20 megawatts and directs the state's environmental agency to assess electricity, water, and land use impacts. Companies must also fund public hearings three months before approval. Democratic Governor Kathy Hochul, who has until December to decide, has not indicated whether she will sign or veto the bill.

Google has released Quantization-Aware Training (QAT) checkpoints for its Gemma 4 model family, optimizing them for on-device use on mobile phones, laptops, and consumer GPUs. Unlike standard post-training quantization, QAT integrates compression into the training process to preserve model quality. A custom mobile quantization schema reduces the Gemma 4 E2B model's memory footprint to under 1GB. Weights are available on Hugging Face in GGUF and compressed tensor formats, with support for llama.cpp, Ollama, LM Studio, vLLM, and MLX.

OpenAI has launched Lockdown Mode, an optional security setting for ChatGPT designed to protect users handling sensitive data from prompt injection attacks, where malicious instructions hidden on webpages trick AI systems into leaking information. The mode limits certain features, disabling Deep Research and Agent Mode entirely, and restricting internet image loading and file downloads. It is available to all accounts including free-tier users via Settings under Safety and Security. OpenAI is also rolling out an active session manager to monitor account access.