Latest AI News

Microsoft has launched hosted agents in Foundry Agent Service, now in public preview. Each agent session gets its own hypervisor-isolated sandbox with persistent filesystem, per-session Entra Agent ID, scale-to-zero billing, and predictable cold starts. Developers can bring any framework — LangGraph, OpenAI Agents SDK, Claude Agent SDK, or Microsoft Agent Framework v1.0 — and deploy with a single azd deploy command. Agents can also publish directly to Microsoft Teams and M365 Copilot.

Microsoft has made agentic capabilities in Word, Excel, and PowerPoint generally available for Microsoft 365 Copilot and Microsoft 365 Premium subscribers. Copilot can now take multi-step, app-native actions directly in documents, worksheets, and presentations rather than just suggesting steps. Early data shows significant gains: Excel engagement rose 67%, retention 50%, and satisfaction 65%. Word saw a 52% engagement increase, while PowerPoint retention climbed 36%. The features are also available to Microsoft 365 Personal and Family plan users.

Anthropic has published guidance on connecting Claude agents to production systems using the Model Context Protocol (MCP), outlining three integration approaches: direct API calls, CLIs, and MCP. MCP is recommended for production cloud-hosted agents due to its standardized auth, portability across clients like Claude, ChatGPT, and Cursor, and rich semantics. MCP SDKs recently surpassed 300 million monthly downloads. Key patterns include grouping tools by intent, using tool search to cut token usage by 85%, and pairing MCP servers with skills for complex workflows.

Anthropic has launched the Economic Index Survey, a monthly study conducted via its Anthropic Interviewer tool to track how AI is reshaping work and the economy. Each month, a randomly selected group of Claude users with accounts at least two weeks old will be invited to share how AI is changing their tasks, productivity, and hiring expectations. Results will feed into future Anthropic Economic Index reports, complementing a companion study analyzing 81,000 open-ended responses collected in December 2024.

OpenAI has launched ChatGPT for Clinicians, a free tool available to verified U.S. physicians, nurse practitioners, physician assistants, and pharmacists. Built with input from hundreds of physician advisors, it supports documentation, medical research, and care consults, and includes trusted clinical search, CME credit tracking, and optional HIPAA compliance. Physician advisors rated 99.6% of responses safe and accurate across 6,924 test conversations. OpenAI also released HealthBench Professional, an open benchmark for evaluating AI on real clinician tasks.

OpenAI has added WebSocket support to its Responses API, cutting agentic workflow latency by up to 40%. Instead of rebuilding full conversation state on each HTTP request, WebSocket mode maintains a persistent connection with in-memory caching of previous response state, tool definitions, and rendered tokens. The update also included safety stack improvements and reduced network hops, achieving a 45% faster time to first token. Codex, Vercel's AI SDK, Cline, and Cursor all reported significant speed gains, with GPT-5.Codex-Spark hitting 1,000 tokens per second.

xAI has launched Grok Voice Think Fast 1.0, its new flagship voice agent model, now available via API. Built for complex, multi-step workflows in customer support, sales, and enterprise applications, it features real-time background reasoning with zero added latency, precise data entry, and support for plus languages. The model tops the τ-voice Bench leaderboard and already powers Starlink's phone support, achieving a 70% autonomous resolution rate and 20% sales conversion rate using 28 distinct tools.

Google unveiled its eighth-generation TPUs at Google Cloud Next, introducing two purpose-built chips: TPU 8t for training and TPU 8i for inference. Designed with Google DeepMind, TPU 8t scales to 9,600 chips delivering 121 ExaFlops per superpod with nearly 3x compute performance over the prior generation. TPU 8i offers 80% better performance-per-dollar and 288 GB of high-bandwidth memory for low-latency agentic workloads. Both chips target the demands of AI agents and will reach general availability later this year.

Google has made Gemini Embedding 2 generally available via the Gemini API and Vertex AI. The natively multimodal embedding model supports search and reasoning across text, image, video, and audio data, eliminating the need for complex, fragmented pipelines. During its preview phase, developers built prototypes including advanced e-commerce discovery engines and video analysis tools. The general availability release brings production-ready stability and optimizations, and the technology also powers multiple Google products internally.

Ideogram has launched Custom Models, a new feature that lets users train image generation models on their own images rather than general internet data. The tool is designed to ensure every generated image follows a brand's specific art direction, typography, and visual identity. Custom Models are available now on Ideogram's Team, Pro, and Enterprise plans and can be accessed via API, addressing a key gap in existing image models that struggle to produce consistent on-brand visuals.

OpenAI has released Privacy Filter, an open-weight model for detecting and redacting personally identifiable information in text. The 1.5B-parameter model, with only 50M active parameters, runs locally to keep unfiltered data on-device. It achieves a 97.43% F1 score on the corrected PII-Masking-300k benchmark and supports up to 128,000 tokens of context. It detects eight PII categories including names, addresses, emails, phone numbers, and API keys. Available under Apache 2.0 on Hugging Face and GitHub.

Alibaba's Qwen team has released Qwen3.6-27B, a dense open-source model designed for coding tasks. Despite having only 27 billion parameters, the model claims to outperform the much larger Qwen3.5-397B-A17B across all major coding benchmarks, demonstrating strong agentic coding capabilities. The release highlights how a smaller dense model can exceed a significantly larger mixture-of-experts model in specialized performance, suggesting improved efficiency in Alibaba's latest generation of open-source AI.

YouTube is expanding its likeness detection tool to the entertainment industry, including talent agencies, management companies, and celebrities. Similar to Content ID, the tool scans for AI-generated deepfakes of a person's face and lets them request removal. Developed with support from major agencies including CAA, UTA, WME, and Untitled Management, the tool is now accessible to celebrities and entertainers even if they do not have a YouTube channel.

Meta is installing tracking software called Model Capability Initiative (MCI) on U.S. employees' computers to capture mouse movements, clicks, keystrokes, and occasional screenshots for AI model training. CTO Andrew Bosworth announced the effort under a rebranded "Agent Transformation Accelerator" program, aiming to build autonomous AI agents that handle work tasks. Meta spokesperson Andy Stone said data won't be used for performance reviews. Legal experts warn the practice may violate European GDPR rules and mirrors surveillance previously limited to gig workers.

X has launched Custom Timelines, one of its biggest changes to the platform's core feed experience. The new feature, powered by Grok, lets users pin specific topics directly to their home tab. With support for over 75 available topics, users can curate a focused feed around their favorite niches. The update gives users significantly more control over what content appears on their main timeline, moving away from a purely algorithmic approach to content discovery.

SpaceX has confirmed a deal giving it the option to acquire AI coding platform Cursor for $60 billion or pay a $10 billion fee for their collaboration. The New York Times first reported the arrangement, which SpaceX confirmed on X. The two companies will work together using SpaceX's Colossus supercomputer, equivalent to one million H100 GPUs. The deal comes ahead of a SpaceX IPO, with Cursor having recently sought $2 billion in funding at a $50 billion valuation.

Google is expanding benefits for AI Pro and Ultra subscribers in AI Studio, offering increased usage limits and access to new models including Nano Banana Pro and Gemini Pro. The update, announced by Product Manager Seth Odoom, lets developers move from idea to working application quickly with predictable costs. For users who've hit free-tier limits, a Google AI plan acts as a low-setup billing bridge before transitioning to pay-per-request API keys for production launches.

Google has added Continued Conversation to Gemini for Home, fulfilling one of users' top requests since early access launched. After an initial "Hey Google" command, the microphone stays active for a few seconds so users can follow up without repeating themselves. Key improvements over Google Assistant include conversational context memory, multilingual support across all supported languages and regions, smarter side-talk detection to reduce accidental responses, and whole-home access for all household members including guests.

Google has open-sourced the DESIGN.md specification from its Stitch AI UI design tool, allowing design rules to be shared across any platform or tool. DESIGN.md lets users export and import design systems between projects, so AI agents understand the reasoning behind color choices and can validate decisions against WCAG accessibility standards. Previously limited to Stitch, the open draft spec is now available on GitHub, enabling broader adoption and consistent AI-generated interfaces that match brand guidelines without rebuilding rules from scratch.

Google has launched Deep Research and Deep Research Max, two next-generation autonomous research agents built on Gemini 3.1 Pro, available via the Gemini API Interactions API in public preview. Deep Research is optimized for speed and low latency, while Deep Research Max uses extended test-time compute for comprehensive, asynchronous workflows. Both agents now support MCP integration for proprietary data sources, native chart and infographic generation, and multimodal inputs. Google is collaborating with FactSet, S&P Global, and PitchBook on MCP server designs.

"Anthropic's Claude has introduced live artifacts inside its Cowork feature, allowing users to build interactive dashboards and trackers that connect directly to their apps and files. These artifacts automatically refresh with current data each time they are opened, removing the need to manually rebuild or update reports. The addition streamlines ongoing data monitoring by keeping outputs continuously synced with live sources, making Claude a more practical tool for users managing recurring workflows and real-time information."

OpenAI has expanded its memory feature in Codex with a new capability called Chronicle, which uses recent screen context to improve memory recall. The update means Codex can now assist users based on what they have recently been working on without requiring them to restate context manually. The memory feature in Codex was initially released as a preview the previous week, and Chronicle represents the next step in that ongoing experiment.

Kimi.ai has released Kimi K2.6, an open-source coding model claiming state-of-the-art results across several benchmarks. The model scores 54.0 on HLE with tools, 58.6 on SWE-Bench Pro, 76.7 on SWE-bench Multilingual, 83.2 on BrowseComp, and 93.2 on Math Vision with Python. A key feature is long-horizon coding capability supporting 4,000 or more steps, positioning K2.6 as a strong open-source option for complex, multi-step programming tasks.

Alibaba's Qwen team has released Qwen3.Max-Preview, an early preview of its next flagship AI model. The release highlights three key improvements over the existing Qwen3.Plus model: enhanced agentic coding capabilities, stronger world knowledge and instruction following, and improved real-world agent and knowledge reliability performance. The preview signals Alibaba's continued push to advance its Qwen model lineup with more capable and reliable AI agents for complex, real-world tasks.

Apple announced that Tim Cook will become executive chairman of its board of directors on September 1, 2026, with John Ternus, currently senior vice president of Hardware Engineering, succeeding him as CEO. The transition was unanimously approved by the board. Ternus, who joined Apple in 2001 and has overseen hardware engineering across iPhone, Mac, iPad, and AirPods, will also join the board. Arthur Levinson, Apple's non-executive chairman for 15 years, will become lead independent director.

Honor's autonomous humanoid robot "Lightning" won a Beijing half-marathon on Sunday, finishing in 50 minutes and 26 seconds — beating all 12,000 human competitors and surpassing the human world record of 57:20 set by Jacob Kiplimo. Developed by smartphone company Honor in just one year, Lightning features legs 90–95 cm long to mimic elite runners and liquid cooling technology. Last year's fastest robot took two hours and 40 minutes. The top three finishers were all autonomous Lightning robots.

TwelveLabs unveiled two major products at NAB Show 2026: Pegasus 1.5, an updated video intelligence model, and Rodeo, a new platform. The company also announced an integration with Autodesk, expanding its reach into professional creative workflows. These launches signal TwelveLabs' push to deepen AI-powered video understanding capabilities for media and entertainment industries, where automated video analysis tools are increasingly in demand for content discovery and production workflows.

Bill Peebles, head of OpenAI's Sora video generation tool, is leaving the company after OpenAI discontinued Sora last month. Kevin Weil, formerly chief product officer and most recently VP of AI for Science, is also departing, with Friday being his last day. OpenAI is shifting focus toward coding and enterprise use to avoid "side quests." Weil's research workspace Prism is being sunsetted, with its capabilities folded into the Codex desktop app.

Anthropic has launched Claude for Word, a Microsoft Word integration now available to subscribers on Pro and Max plans. The release coincides with the introduction of Claude Opus 4.7, Anthropic's latest model. The Word integration allows users to access Claude's writing and editing capabilities directly within Microsoft Word documents, bringing AI assistance into one of the world's most widely used word processing applications.

Dairy Queen is rolling out Presto's AI voice chatbot to select drive-thrus across the US and Canada, aiming to speed up service and upsell customers. Following a test last year, the chain is broadly deploying the technology, which correctly takes orders about 90 percent of the time. Presto already works with Carl's Jr., Hardee's, and Taco John's. A 2023 Bloomberg report noted Presto's AI drive-thrus may be partly assisted by human workers in the Philippines.

Netflix is launching a TikTok-style vertical video feed within its apps this month, after testing the feature since 2025. The feed will help users discover video podcasts, shows, and movies. Co-CEO Gregory Peters said AI-powered recommendation systems will improve personalization, while co-CEO Ted Sarandos said GenAI will enhance content creation tools. Netflix also acquired Ben Affleck's AI filmmaking company InterPositive to accelerate its GenAI capabilities. The company reported Q1 2026 revenue of $12.25 billion, up 16.2% year-over-year, and targets $3 billion in ad revenue this year.

Canva has integrated with Anthropic's Claude Design, allowing users to turn AI-generated drafts directly into fully editable designs within Canva's Visual Suite. The partnership also introduces HTML importing, making Canva the first platform to unify visual, document, and interactive content in a single collaborative editor. Users can edit AI-generated HTML artifacts using drag-and-drop without regenerating code. The collaboration builds on two years of work, including the Canva MCP for Claude launched last July, which millions have already used.

HeyGen has open-sourced HyperFrames, an agent-native framework that converts HTML to MP4 video, built using Claude Code. The tool is designed for AI agents to programmatically generate video content and can be installed via the command npx skills add heygen-com/hyperframes. HeyGen used HyperFrames to produce its own launch video entirely within Claude Code, demonstrating the framework's practical capability for automated video production workflows.

Canva has launched AI 2.0, its biggest platform overhaul since moving design to the browser, introducing prompt-based editing tools that let users create or modify designs using plain text descriptions. A new orchestration layer connects all of Canva's tools through a single conversational interface, enabling commands like generating a full multi-channel campaign from one prompt. The update also adds persistent memory for consistent branding, Object-Based Intelligence for precise element editing, and HTML import support. It launches as a research preview for the first one million users.

Opera has launched Browser Connector, a free feature now available in Early Bird mode for Opera One and Opera GX that connects ChatGPT or Claude directly to the browser via Model Context Protocol (MCP). The AI can read open tabs, take screenshots, navigate to pages, and optionally access browsing history. Users retain control over permissions, and the AI cannot autonomously click or fill forms. The feature aims to eliminate constant copy-pasting of context between browser and AI tools.

MZLA Technologies Corporation, a Mozilla Foundation subsidiary, has launched Thunderbolt, an open-source self-hosted AI client designed to give organizations full data ownership and freedom from vendor lock-in. Built with native integration with deepset's Haystack agent and RAG orchestration framework, Thunderbolt supports chat, search, automation, and cross-device workflows on Windows, macOS, Linux, iOS, and Android. CEO Ryan Sipes calls it a "sovereign AI client." The source code is on GitHub, and enterprise deployments are available via waitlist at thunderbolt.io.

Character.ai has launched c.ai Books, an interactive feature that lets users step inside classic public domain novels and roleplay as characters in real time. Available now on mobile and web via c.ai Labs, the library includes 20+ titles sourced from Project Gutenberg, including Alice in Wonderland, Pride and Prejudice, Dracula, and Frankenstein. Users can follow the original plot or go off-script in freeform mode. Free users get limited turns; c.ai+ subscribers receive a higher allowance. Community-created Alternate Universe remixes are also supported.

Google has updated its Gemini app with Personal Intelligence image generation powered by Nano Banana 2 and Google Photos integration. Users can generate personalized images using simple prompts like "design my dream house" without lengthy descriptions or manual photo uploads. Connecting Google Photos lets Gemini use labeled people and pets to create images in styles like watercolor or claymation. The feature is rolling out to U.S. Google AI Plus, Pro, and Ultra subscribers, with Chrome desktop support planned. Google does not train models on private photo libraries.

Perplexity has launched Personal Computer for Mac, an expansion of its Perplexity Computer product that uses multi-model AI orchestration to work across local files, native apps, connectors, and the web. Available today to Perplexity Max subscribers via waitlist, it integrates with macOS tools like Notes, iMessage, and email, and is activated with a dual-CMD shortcut. Files are handled in a secure sandbox with auditable, reversible actions. A Mac mini is recommended for 24/7 persistent access.

OpenAI has launched GPT-Rosalind, a frontier reasoning model purpose-built for life sciences research, drug discovery, and translational medicine. Named after DNA pioneer Rosalind Franklin, the model is optimized for chemistry, protein engineering, and genomics workflows. It outperforms GPT-5.4 on 6 of 11 LABBench2 tasks and ranked above the 95th percentile of human experts on RNA sequence prediction. Partners include Amgen, Moderna, and the Allen Institute. A free Life Sciences plugin for Codex connects to over 50 scientific databases. Access is restricted to qualified U.S. enterprise customers.

OpenAI has expanded its Trusted Access for Cyber program and committed $10 million in API credits to defenders through its Cybersecurity Grant Program. Initial recipients include Socket, Semgrep, Calif, and Trail of Bits, focusing on supply chain security and vulnerability research. Major organizations including Bank of America, Cisco, CrowdStrike, JPMorgan Chase, and Goldman Sachs have joined the program. OpenAI also provided access to GPT-5.Cyber to the U.S. CAISI and UK AI Security Institute for capability evaluations.

Alibaba's Qwen team has open-sourced Qwen3.6-35B-A3B, a sparse Mixture-of-Experts model with 35 billion total parameters but only 3 billion active parameters. Released under the Apache 2.0 license, the model delivers agentic coding performance comparable to models ten times its active size. It also features strong multimodal perception and reasoning capabilities, along with support for both multimodal thinking and non-thinking inference modes.