This week made one thing clear: the way we access, own, and deploy AI is being rewritten. Anthropic redeployed Claude Fable 5 - its most capable public model - globally after the US government lifted export controls, and from July 7 it moves to metered usage, where you pay for every prompt beyond your plan's limits rather than a flat subscription. Alongside it, the company launched Claude Science, a research workbench with 60+ scientific databases and a citation-checking agent, and made the cheaper Sonnet 5 its new default model at $2 per million input tokens.

Microsoft took a different route to ownership, unveiling the Microsoft Frontier Company - a new unit committing $2.5 billion and roughly 6,000 engineers to embed inside enterprises like the London Stock Exchange Group, Unilever, and Accenture, co-building AI systems those companies own outright. Customers can run any model underneath, from OpenAI to open-source, and Microsoft promises their data won't train its models. Meanwhile, physical AI reached the factory floor: Figure's third-generation humanoid, F.03, arrived at BMW's Plant Spartanburg to handle parts sequencing, graduating from the body shop after its predecessor supported production of more than 30,000 X3s.

The race to own the entire AI stack - chips, models, and distribution - defined this week. OpenAI unveiled Jalapeño, its first custom AI chip, built with Broadcom and designed purely for inference. Taken from design to tape-out in nine months, it promises performance-per-watt substantially better than today's best, with deployment targeted for late 2026. The company also shipped GPT-5.5-Cyber, a frontier cybersecurity model gated to vetted defenders like Cisco and CrowdStrike, and struck a Getty licensing deal that puts over 400 million real editorial images inside ChatGPT search.

Anthropic pushed collaboration in a different direction with Claude Tag, a beta that drops a single shared Claude into Slack as a team member with its own identity and memory, summoned by tagging @Claude. Rather than a private assistant, one Claude serves a whole channel - following context over time, chasing quiet threads, and running long tasks over hours or days, all scoped and logged by admins. Geopolitics loomed over it all: Beijing hit 56 American companies with trade curbs, placing rare-earth miners and drone makers under full export controls in direct response to the Pentagon's expanded blacklist of Chinese military-linked firms.

OpenAI has officially launched GPT-5, the most capable AI model the company has ever built. The new model brings major advances in multi-step reasoning, coding assistance, and scientific problem-solving, while significantly reducing the hallucination rate that plagued earlier versions.

CEO Sam Altman described GPT-5 as a defining moment in AI development. Enterprise API access goes live today, with broader rollout expected within weeks.

AI spilled into new territory this week - from rockets to radiology. Days after its record $75 billion IPO, SpaceX agreed to buy AI coding tool Cursor for $60 billion in an all-stock deal, roughly double Cursor's valuation from November. The move folds the most-used coding tool in enterprise straight into xAI's Colossus supercomputer, and Cursor's CEO promptly unveiled a 1.5-trillion-parameter model trained on more than 100,000 GPUs. Rather than spend years closing the gap on rival models, SpaceX simply bought the tool engineers already open every day.

OpenAI, meanwhile, taught its agents to learn by watching. Its new Record & Replay feature for the Codex app observes you complete a task once - filing an expense, pulling a weekly report - then turns that demonstration into a reusable, editable skill. Paired with rebuilt scheduled tasks that ping you only when something meaningful changes, it's an agent that learns your routine once and runs it on a timer. And in the week's most unexpected pivot, image company Midjourney launched a medical division and the Midjourney Scanner, a full-body ultrasound that images your entire body in roughly 60 seconds with no radiation, targeting a billion scans a month by 2031.

This week, AI stopped being just a cloud-based tool and started living on our desktops, in our workflows, and even in our hardware. Microsoft kicked things off at Build 2026 by unveiling a complete in-house AI ecosystem: seven new models, including the reasoning-focused MAI-Thinking-1, an always-on work agent called Scout, and Project Solara, a hardware platform designed for agent-first devices. Meanwhile, NVIDIA introduced RTX Spark, a chip powerful enough to run advanced AI agents directly on Windows laptops, and OpenAI rolled out updates that make ChatGPT more proactive, specialized, and integrated into daily tasks than ever before.

Microsoft's move signals a shift toward self-sufficiency, offering its own models, agents, and hardware while keeping its platform open to third-party tools. NVIDIA's RTX Spark, packed with a petaflop of AI compute, aims to bring agentic AI to laptops and desktops this autumn, challenging the cloud-first status quo. OpenAI, on the other hand, is focusing on ownership - its new Dreaming system automatically updates ChatGPT's memories, and role-specific Codex agents are now available for non-coders, making AI a seamless part of workflows across industries. The message is clear: AI's next phase is about integration, not just innovation.

This week showed both the promise and the fragility of frontier AI. Anthropic launched Claude Fable 5 on June 9 - the most capable model it had ever released for general use, priced to spread at less than half of Mythos Preview and free across paid plans. Three days later, the US government invoked export controls and ordered it switched off for every customer worldwide, citing a reported jailbreak, with no published safety finding and no wind-down for those who had just paid to upgrade. A utility you can be unplugged from by government order, it turns out, is a very different kind of utility.

Apple, meanwhile, made peace with renting. At WWDC it introduced Siri AI, a ground-up rebuild of its assistant with Google's Gemini models doing the heavy lifting in the cloud - a reported billion-dollar-a-year deal that ends a decade of Apple insisting it would do AI on its own terms. iOS 27 will even let users set Claude or ChatGPT as their default assistant, though conflicts with the Digital Markets Act mean it won't ship in the EU on day one. And SpaceX priced the largest IPO in history, raising around $75 billion at a valuation near $1.75 trillion, folding xAI and Grok into the listing.

Microsoft used its Build 2026 developer conference in San Francisco to unveil a full in-house AI stack: seven first-party MAI models led by MAI-Thinking-1, a 35-billion-parameter reasoning model trained from scratch, an always-on 'Autopilot' work agent called Scout that handles meeting prep and scheduling across Teams and Outlook, and Project Solara, a platform for agent-first hardware that debuted with a Qualcomm wearable badge and a MediaTek desk hub. After years of leaning on partners for its frontier models, Microsoft now has a complete stack of its own - while deliberately keeping the platform open to Claude and OpenAI in the same workflow.

NVIDIA answered at its GTC Taipei keynote with RTX Spark, a chip built to run AI agents directly on a Windows laptop or desktop rather than in the cloud. Pairing a Blackwell RTX GPU with a 20-core Arm CPU and up to 128GB of unified memory, it reaches roughly a petaflop of compute - enough to run a 120-billion-parameter model locally - and ships this autumn across more than 30 laptops from Dell, HP, ASUS, Lenovo and MSI. OpenAI, meanwhile, widened who it builds for: a new 'Dreaming' system keeps ChatGPT's memories current, six Codex role plugins turn it into a specialist for non-coders, and an upgraded GPT-Rosalind model series targets life sciences and drug discovery.

The numbers around AI reached staggering new heights this week. Anthropic upgraded its flagship to Claude Opus 4.8 - a sharper, more honest model that its own evaluations show is around four times less likely to let flaws in its own code slip past - at the same price as 4.7. On the same day, the company raised $65 billion in Series H funding at a $965 billion post-money valuation, edging toward the trillion-dollar mark, with run-rate revenue that crossed $47 billion earlier in the month. Alongside the model, Claude Code gained dynamic workflows that can spin up hundreds of parallel subagents to carry codebase-scale migrations from kickoff to merge.

Robotics and space matched the ambition. Figure signed a commercial agreement with Catalyst Brands - the retail group behind JCPenney, Aéropostale, and Brooks Brothers - to deploy its humanoids at scale, starting in a Reno distribution centre, marking its move from a single BMW car line into general retail logistics. And SpaceX filed its IPO prospectus with the SEC, applying to list on Nasdaq under the ticker SPCX in what would become the largest stock-market debut in history, with the listing housing Musk's entire AI operation after xAI and Grok were folded in earlier this year.

At I/O 2026, Google declared the 'agentic Gemini era' - a shift from assistants that answer prompts to agents that take action across its products - and rebuilt its entire line-up around it. Gemini 3.5 Flash is now generally available and beats last year's Pro model on agentic benchmarks while running four times faster at under half the cost of rival frontier models. Gemini Omni turns any text, image, audio, or video into editable video with real-world physics understanding, which Demis Hassabis called a step toward AGI, and AI Mode in Search - now past a billion monthly users - got its biggest redesign in 25 years.

The week also revealed AI's hardware ambitions and its operational limits. Brett Adcock's secretive lab Hark raised over $700 million at a $6 billion valuation - before shipping a single product - with Nvidia, AMD, Intel, and Qualcomm all joining the round to build a 'personal intelligence' on a new class of device. Yet reality bit elsewhere: Starbucks retired its AI inventory tool just nine months after rolling it out across 11,000 stores, sending baristas back to counting milk by hand after the system, sold on 99% accuracy, frequently miscounted in busy backrooms.

Google spent the days before I/O 2026 dropping update after update, and the bigger picture is starting to become obvious. Instead of treating Gemini like a separate chatbot, Google is turning it into an AI layer that sits underneath everything — Android, laptops, apps, and even how users interact with their devices.

One of the most interesting launches is Gemini Intelligence, which can handle tasks across apps without constantly switching between them. Google also introduced “Magic Pointer,” a new way of interacting with Gemini through gestures, motion, and voice instead of traditional clicks. At the same time, DeepMind spinout Isomorphic Labs secured a massive funding round to continue its AI-driven drug discovery work, showing how Google’s AI ecosystem is expanding far beyond consumer products into healthcare and scientific research.

The battle for enterprise AI intensified this week, with Anthropic pulling ahead on several fronts. The company disclosed that Q1 2026 revenue grew 80x year-on-year to an estimated $44 billion-plus run rate, committed $200 million to a four-year Gates Foundation partnership across global health and education, and overtook OpenAI in verified business customers for the first time - reaching 34.4% versus OpenAI's 32.3%. Customers spending over $1 million annually climbed from a dozen two years ago to more than a thousand today.

OpenAI answered by mirroring Anthropic's biggest enterprise plays within a single week. It launched a Deployment Company with $4 billion from 19 investment and consulting firms, acquired UK consultancy Tomoro for its forward-deployed engineers, and shipped Daybreak, a cyber-defensive initiative built on GPT-5.5-Cyber with vendors like Cloudflare and CrowdStrike. Google, never one to stay quiet before a keynote, flooded the days before I/O with updates: Gemini Intelligence as an agentic layer under Android, DeepMind's Magic Pointer reimagining the mouse cursor, and a $2.1 billion Series B for Isomorphic Labs to chase its mission to 'solve all disease.'

SpaceXAI just made one of the biggest infrastructure deals in AI so far. After SpaceX absorbed xAI earlier this year, the newly combined company signed a massive agreement with Anthropic that basically hands over the entire compute capacity of the Colossus 1 data center to power Claude. That means hundreds of thousands of NVIDIA GPUs and an enormous amount of energy capacity are now focused almost entirely on AI inference.

The move also explains why Anthropic suddenly increased usage limits across Claude products. With far more compute available, the company can support bigger workloads without the same bottlenecks as before. At the same time, the deal is financially smart for SpaceXAI because Colossus 1 was no longer their main training cluster after shifting to the even larger Colossus 2 system. Instead of leaving that infrastructure underused, they’re turning it into a huge source of recurring revenue while continuing to expand their AI ambitions.

OpenAI spent the week reshaping how its technology gets distributed and integrated across the cloud industry. The biggest shift came from changes to its partnership with Microsoft, which now gives OpenAI much more freedom to sell its models across different cloud providers instead of being tied mainly to Azure. That’s a pretty major move because it opens the door for OpenAI to become available almost everywhere enterprises already operate.

At the same time, OpenAI expanded aggressively into Amazon’s ecosystem by bringing its frontier models and managed agents to AWS Bedrock. On top of that, the company open-sourced a system called Symphony, designed to coordinate autonomous AI agents through tools like issue trackers and development workflows. Altogether, the updates show OpenAI moving beyond just building models and focusing more on becoming the infrastructure layer behind future AI-powered work.

Anthropic spent this week pushing Claude far beyond being just a chatbot. Instead of focusing on one big flashy release, they rolled out a bunch of upgrades that make Claude more connected to real workflows, developer tools, and even everyday consumer apps. The overall direction is pretty obvious: Claude is turning into something that lives across your whole digital environment instead of sitting in a single chat window.

One major addition is live artifacts inside Cowork, where Claude can build dashboards and trackers that stay connected to apps and files with constantly updated data. Developers also gained a new memory system for managed agents, allowing AI systems to remember information across sessions in a way that’s editable, exportable, and easier to control. On top of that, Claude now integrates with a huge list of consumer services like Spotify, Uber, Booking.com, and Instacart, making it much more useful outside of work-focused tasks.

Google is rolling out Gemini updates at an almost ridiculous speed right now. Between a new top-performing text-to-speech model and the launch of a native Mac app, it’s obvious they want Gemini to be everywhere at once — from developer tools to everyday desktop use.

One of the biggest launches is Gemini 3.1 Flash TTS, which now ranks at the top of major text-to-speech leaderboards. The system can control tone, pacing, accents, and emotions directly through prompts, making AI voices sound way more natural and expressive. At the same time, Google finally released a native macOS app for Gemini with live screen sharing and file context support, which makes the assistant feel much more integrated into actual workflows instead of just being another browser tab.

Some of the biggest names in AI — OpenAI, Anthropic, and Google — are starting to collaborate in a way that would’ve seemed unlikely not long ago. Through the Frontier Model Forum, they’ve begun sharing information to detect and respond to competitors trying to extract capabilities from their models. The focus is on protecting their most advanced systems from being copied or reverse-engineered.

A major concern is something called adversarial distillation, where outputs from powerful models are used to train smaller, cheaper versions. According to reports, this has already cost US AI companies billions in potential revenue. Anthropic even tracked millions of suspicious interactions and thousands of fake accounts linked to specific organizations. What makes this situation stand out is that this isn’t just discussion anymore — it’s an active intelligence-sharing effort between companies that are usually direct competitors.

Google just had one of its biggest AI release weeks in a while, and the direction is pretty clear: more openness and more real-world integration. With the launch of new open models and updates to everyday tools like Gmail, the company is trying to reach both developers and regular users at the same time.

One major step is the release of Gemma 4, a set of open models built on the same research as Gemini 3. These models range from lightweight versions that can run directly on devices to larger ones that compete surprisingly well with much bigger systems. On top of that, Google is making them easier to use commercially by removing previous licensing restrictions. At the same time, the new AI-powered Gmail inbox shows how this tech is moving into daily workflows, replacing simple unread counters with summaries, priorities, and task suggestions generated directly from your emails.

Brett Adcock, known for founding Figure AI, is back with a new project called Hark, and he's going all in. Backed by $100 million of his own money, the goal is to build what he calls "the most advanced personal intelligence in the world." Instead of launching a single product, Hark is aiming to build an entire AI system from scratch.

What makes Hark stand out is how broad the approach is. Rather than focusing on just software or models, the team is working on everything at once — foundation models, software, and custom-built hardware, all designed to work together as one system. The team itself is stacked, with talent coming from companies like Apple, Google, Tesla, and Amazon. On top of that, the timeline is aggressive, with plans to deploy large-scale GPU infrastructure soon and release their first models within months, followed by dedicated hardware shortly after.

Anthropic just dropped a bunch of new features in a really short time, and it honestly feels like they're trying to catch up fast. With updates like Dispatch, Claude Code Channels, and Projects in Cowork, they're basically building their own version of an AI agent system that can handle real tasks across devices. Instead of just chatting with Claude, the focus is clearly shifting toward letting it actually do work in the background.

One of the most interesting parts is how connected everything is becoming. You can message Claude from your phone while it works on your desktop, assign tasks like organizing files or writing reports, and then come back later to finished results. Developers can even interact with Claude Code through apps like Telegram or Discord, which makes it feel more like texting a coworker than using a tool. On top of that, Projects brings everything into one place, keeping files, instructions, and context organized locally, which also adds a layer of privacy.

Yann LeCun is starting fresh with a new company called AMI Labs, and the scale of it is already huge. They secured over $1 billion in funding right away, even though there's no product yet. Instead of building another chatbot, the goal is to rethink how AI systems work at a much deeper level.

The focus is on something called "world models." Instead of predicting the next word like most current AI systems, these models aim to understand how things change over time and how different factors influence each other. That would make AI less about generating text and more about actually understanding cause and effect. Even without a clear business model yet, major players like NVIDIA are backing the idea, which shows how much confidence there is in this direction.

OpenAI just released GPT-5.4, and it looks like they're trying to push their models further into real professional work. Instead of focusing on one specific ability, GPT-5.4 combines reasoning, coding, and agent-style workflows into a single system designed to handle complex tasks from start to finish. The idea is basically that AI shouldn't just answer questions anymore — it should be able to actually carry out work.

Early benchmarks suggest it's already performing extremely well across many knowledge-work tasks. In tests covering dozens of professions, GPT-5.4 matched or even beat human experts most of the time. Another big step is that the model can now operate computers directly, navigating software and completing actions in real environments instead of just generating instructions. On top of that, improved tool handling and a massive context window mean the system can plan longer workflows and use external tools more efficiently without wasting tokens.

Google launched Nano Banana 2 (NB2), a major upgrade to its image generation model that is now the default across Gemini and AI Studio. The killer feature? It grounds itself in Google Search to generate factually accurate visual content, like properly labelled cross-sections of engines.

They've also largely solved the text rendering problem. NB2 hits around 95% accuracy for text in images, minimizing the garbled fonts that usually plague AI art. Add in full aspect ratio control up to 4K and character consistency, and it's a massive leap forward.

Anthropic had a wild week. CEO Dario Amodei publicly refused the US Department of War's demand to remove safeguards against autonomous weapons and mass surveillance, stating the tech isn't reliable enough and that surveillance breaks democratic values.

On the product side, they updated their Cowork desktop agent with private plugin marketplaces and deep integrations for finance, allowing Claude to work across Excel and PowerPoint in a single session. Meanwhile, Claude Code hit a $2.5 billion run rate, launching a remote control feature for mobile.

Perplexity launched "Computer," a cloud-based AI system that orchestrates 19+ frontier models into a single autonomous workflow engine. Users describe a high-level goal, and the system breaks it down—assigning Claude for reasoning, Gemini for research, and Grok for speed.

It operates asynchronously in the background for hours or months, checking in only when it needs human input. Currently priced at $200/month for Max subscribers, it's being positioned as a managed alternative to complex open-source setups.

Perplexity just launched something called Computer, and the idea behind it is pretty ambitious. Instead of using just one AI model for everything, Computer connects more than 19 different frontier models into a single workflow system. The goal is that you describe what you want to achieve, and the system automatically breaks the task down and assigns each step to the model that’s best at it.

For example, one model might handle reasoning, another might do deep research, while others generate images, videos, or quick search results. What makes it interesting is that it doesn’t just respond to a single prompt like normal chatbots. Computer can run tasks in the background for hours or even days, managing full projects and only asking the user for input when it actually needs it. Right now it’s available to Perplexity Max subscribers, but the bigger idea is clearly to turn AI into something that manages complex workflows almost like a digital operating system.

Anthropic dropped Claude Sonnet 4.6, and it's starting to cannibalize their flagship model. Early testers actually prefer it over Opus 4.5 in 59% of coding tasks. It costs a fraction of Opus ($3/$15 per million tokens) but delivers better instruction following and fewer hallucinations.

It's also crushing computer use benchmarks. On OSWorld, it's hitting human-level navigation across complex spreadsheets and web forms. Plus, with a 1M token context window, it can hold entire codebases in memory at once, making it incredibly capable for long-horizon planning.

Google shipped three massive updates in a single week. First, Gemini 3.1 Pro more than doubled its reasoning score on the ARC-AGI-2 benchmark, hitting 77.1%. Second, DeepMind's Lyria 3 is now generating 30-second music tracks with vocals directly inside the Gemini app, complete with SynthID watermarks.

But the sleeper hit is Pomelli, a free Google Labs tool. You upload a basic, poorly lit phone photo of a product, and it spits out professional, studio-quality lifestyle imagery for free—no photographer needed.

David Silver, the mind behind DeepMind's AlphaGo, is raising $1 billion in seed funding for his new London-based lab, Ineffable Intelligence. It's the largest first round in European startup history, drawing interest from Nvidia, Google, and Microsoft.

Silver is betting heavy on reinforcement learning—training systems through experience and environmental interaction rather than just scraping static internet text. He's part of a massive wave of top scientists, including Ilya Sutskever and Mira Murati, leaving Big Tech to start their own ventures.

Anthropic just released Claude Sonnet 4.6, and it’s a surprisingly big upgrade for what’s supposed to be their “mid-tier” model. Early testers are saying it performs almost like the much more expensive Opus model, especially in things like coding, reasoning, and complex tasks. What makes it interesting is that it manages to get that performance without the same cost, which could make it a much more practical option for developers and companies.

One of the biggest improvements is how well it handles real computer tasks. Users reported that Sonnet 4.6 can navigate complicated spreadsheets, fill out multi-step web forms, and work across multiple browser tabs in ways that feel almost human. On top of that, the model now supports a 1-million-token context window, meaning it can analyze huge amounts of information at once, like entire codebases or large research documents. That makes it much better for long-term reasoning and complex workflows.

ElevenLabs just held its London Summit following a massive $500M Series D led by Sequoia. That brings their valuation to $11B, tripling from a year ago. They closed 2025 with over $330M in ARR, pulling in enterprise heavyweights like Deutsche Telekom and Revolut.

The tech is getting serious. Their new Expressive Mode makes voice agents emotionally intelligent—adjusting tone in real-time to de-escalate frustrated callers. They're also pushing hard into government. In the Czech Republic, AI agents are already handling 5,000 calls a day with an 85% independent resolution rate.

Anthropic just closed a $30 billion Series G round, pushing its valuation to an eye-watering $380 billion. They're now the second most valuable AI lab globally, right behind OpenAI. The numbers back it up: $14 billion in annualized run-rate revenue, growing 10x annually for three straight years. Eight of the Fortune 10 are already using Claude.

A massive driver here is Claude Code, which just crossed a $2.5 billion run-rate. It's wild to think an estimated 4% of all public GitHub commits are now authored by Claude. Add the fact that Anthropic is the only frontier model available across AWS, Google Cloud, and Azure, and their enterprise reach is essentially unmatched right now.

Chinese tech giant ByteDance launched Seedance 2.0, a video generation model that produces cinematic-quality clips from basically any input, immediately triggering a firestorm from Hollywood over intellectual property theft. The multimodal architecture delivers director-level control over lighting and camera movement, leading internal benchmarks across the board.

Disney fired back instantly with a cease-and-desist letter, accusing ByteDance of pre-packaging Seedance with pirated characters from Star Wars and Marvel. The wider industry is equally furious. The Motion Picture Association and SAG-AFTRA condemned the use of actors' likenesses without consent, calling it a "virtual smash-and-grab."

ElevenLabs is starting to look less like a niche AI startup and more like a major player in the voice space. After raising $500 million and hitting an $11B valuation, they’re now pushing hard on making voice agents actually usable in real-world scenarios. The focus isn’t just on sounding realistic anymore, but on making conversations feel natural and responsive.

A big part of that is their new “Expressive Mode,” which allows voice agents to pick up on emotions like stress and adjust how they speak in real time. That means conversations can feel less robotic and more human, especially in situations like customer support. At the same time, ElevenLabs is expanding into the public sector, where their agents are already handling thousands of calls per day and solving most of them without human help. It’s a clear sign that voice AI is moving from demo to real deployment.

Elon Musk just merged SpaceX and xAI in a massive $1.25 trillion deal, creating what could be the most valuable private company ever. The move doesn't just combine rockets and AI, it also sets up a potential blockbuster IPO that some estimate could be worth around $50 billion. It's one of those announcements that sounds almost unreal, even by Musk standards.

The idea behind the merger is big and kind of wild. Musk wants to push AI infrastructure into space, arguing that Earth's power grids won't be able to handle AI's future energy needs. SpaceX has already asked regulators for permission to massively expand Starlink into an "orbital data center system," jumping from around 9,400 satellites today to potentially over a million. At the same time, critics point out that xAI is burning huge amounts of money and still trails competitors like OpenAI and Google, making some people see this deal as risky financial engineering rather than pure innovation.

Google DeepMind's Project Genie lets users create and explore interactive worlds using just text prompts and images. Instead of generating a single static scene, Genie builds the environment ahead of you in real time as you move through it, which makes the experience feel much more alive. Even though it's still a research prototype, the idea alone already feels like a big shift.

What makes Genie especially interesting is that it goes far beyond gaming. The system can simulate physics and interactions in a way that could be useful for robotics, animation, training simulations, or exploring historical and fictional environments. On top of that, Gemini's new Agentic Vision turns image understanding into an active process, where the AI can zoom in, inspect, and manipulate visuals step by step instead of just analyzing them once.

Anthropic is clearly pushing Claude beyond being just a chatbot and more into a real productivity tool. By expanding Claude in Excel to Pro plans, a lot more users can now use it for actual spreadsheet work instead of just testing it in limited environments. The focus seems to be on making Claude fit naturally into everyday workflows, not just sit in a separate AI interface.

At the same time, Anthropic is experimenting with health data connections, allowing Claude to summarize and explain medical information when users explicitly opt in. Alongside these practical updates, they also published the full Constitution that defines how Claude should behave, outlining priorities like safety, ethics, and helpfulness. It's a pretty transparent move that shows how seriously they're taking the idea of AI having a defined "character."

Anthropic released Cowork, a research preview that brings Claude Code’s agent-style abilities directly to the Claude Desktop app. Instead of just answering questions, Claude can now actually work through tasks on its own, making it feel more like a digital coworker than a chatbot.

Cowork allows users to describe a goal in plain language and then let Claude figure out the steps needed to get there. It can directly read and write local files, meaning it can create proper Excel sheets with formulas, PowerPoint presentations, and well-formatted documents without manual copying. For more complex tasks, Claude splits the work into smaller subtasks and runs them in parallel, handling things like research, data processing, and synthesis almost completely on its own.

NVIDIA used CES to clearly show where their focus is going next: physical AI. Instead of just text and images, NVIDIA is pushing AI into the real world, like robotics, self-driving cars, and systems that actually interact with physical environments. This makes AI feel a lot less abstract and way more impactful.

One big step is AlpamayO, NVIDIA’s new open-source model for autonomous driving. It doesn’t just make decisions but explains them step by step, which is huge for safety and trust. On top of that, NVIDIA introduced Cosmos, a set of simulation models that let developers train autonomous systems in virtual worlds before they ever hit real roads. Finally, the new Rubin AI platform shows that NVIDIA isn’t just doing research — they’re scaling this tech for real production, with much cheaper and more efficient hardware coming soon.

Meta announced that it bought an AI startup called Manus for over $1 billion, which is a huge move in the AI race. Manus focuses on building autonomous AI agents that can plan, use tools, and execute tasks on their own instead of just answering questions. Meta plans to integrate this system into products like WhatsApp, Messenger, and even smart glasses. What makes this deal interesting is that Manus didn’t win by having a better AI model, but by building a smarter structure around the model.

This shows a bigger trend in AI: raw intelligence isn’t everything anymore. How AI is deployed, connected to tools, and scaled across products matters just as much. By buying Manus, Meta skips years of internal development and gets a system that already works in real-world scenarios. It also positions Meta strongly for future AI assistants that actually act instead of just chatting.

Nvidia and Stanford University released NitroGen, an open-source AI that can play more than 1,000 video games. Instead of learning one game at a time, it was trained on around 40,000 hours of gameplay videos from YouTube and Twitch. By watching humans play, NitroGen learned controls, strategies, and reactions. What’s impressive is that it also performs well in games it has never seen before.

This shows real progress toward general gaming AI instead of game-specific bots. Because NitroGen is open-source, developers and researchers can improve it freely. That could speed up innovation in gaming AI a lot.

OpenAI released several updates that improve how ChatGPT handles images, coding, and daily use. With GPT Image 1.5, images are generated faster, look sharper, and handle lighting, details, and even text much better. OpenAI also upgraded Codex, making it stronger for long and complex programming tasks like refactoring. On top of that, ChatGPT got usability features like writing blocks, pinned chats, and personalization options.

These changes might not sound dramatic, but they make ChatGPT feel more polished and reliable. Instead of focusing on big promises, OpenAI is improving the small things people use every day. This makes the tool more practical for school, work, and creative projects.

Disney announced that it’s investing $1 billion into OpenAI and signing a three-year licensing deal that lets people create AI-generated videos and images using over 200 characters from Disney, Pixar, Marvel, and Star Wars. These creations will be made using tools like Sora and ChatGPT Images, which shows that Disney is no longer just watching AI from the sidelines. What makes this even crazier is that just a day earlier, Disney had sent Google a cease-and-desist letter over large-scale copyright issues.

This move shows a big strategy change: instead of fighting AI everywhere, Disney is choosing to license its content where it makes sense. One major issue with AI is that it can basically memorize famous characters, which creates legal risks—often called the “Snoopy problem.” By licensing its characters, Disney turns a legal headache into something officially allowed. On top of that, Disney is becoming a premium data partner at a time when AI companies are running out of high-quality training material. Its huge character library is now a powerful asset, not just something to protect.

ByteDance released a new AI video model called Vidi2, and it’s beating some of the strongest AI systems on video understanding benchmarks. The model is especially good at understanding what’s happening across time in videos, finding specific moments, and answering questions about video content. What makes this impressive is that it combines several skills—like tracking objects, understanding scenes, and answering questions—into one system instead of separate tools.

Vidi2 outperformed competing models on multiple benchmarks, especially when it comes to understanding motion across frames and quickly finding very short video moments. It can handle videos ranging from just a few seconds up to half an hour, which makes it useful for real-world applications. Because of this, the model isn’t just for research but also fits professional workflows like video editing, automatic camera switching, and tracking characters across scenes. Overall, it shows how fast video-focused AI is improving.

U.S. President Donald Trump signed an executive order launching the "Genesis Mission," a massive national project designed to speed up scientific discovery using AI. The idea is similar to the Manhattan Project, but instead of weapons, it focuses on science and technology.

The U.S. government wants to combine powerful supercomputers, huge federal datasets, and AI agents to automate research and test scientific ideas faster than humans alone could. This would all run on existing government research infrastructure.

Department of Energy Leadership

The Department of Energy will be in charge of building the platform, connecting national labs, universities, and approved private companies. The plan moves very fast: within a few months, officials must identify major scientific challenges and quickly show real results.

Key Research Areas

These challenges include areas like biotech, nuclear fusion, quantum computing, semiconductors, and advanced manufacturing. Strict cybersecurity rules are meant to protect sensitive research while still allowing collaboration.

Overall, the project shows how seriously the U.S. is taking AI as a strategic tool for science and national security.

ElevenLabs, best known for realistic AI voices, is now expanding into images and video with a new Image & Video platform (currently in beta). Instead of building everything from scratch, ElevenLabs connects top image and video models like Sora, Veo, and Kling into one unified workspace.

The idea is to let creators generate visuals, add AI voiceovers, include music, and layer sound effects all in one place. This turns ElevenLabs from a voice tool into a full content creation platform.

Unified Workflow

What makes this move strong is the workflow focus. Users don't need to jump between different apps for images, video, and audio anymore. Everything happens inside one timeline, from the first idea to the final export.

Targeting Creators

By targeting creators, marketers, and content teams, ElevenLabs is positioning itself as a serious alternative to using multiple separate tools. It's less about having the best single model and more about making creation faster and smoother.

OpenAI released GPT-5.1, an update that makes ChatGPT feel smarter, warmer, and more adaptable depending on the situation. The new version can adjust how much it thinks before answering, so it responds quickly to simple questions but takes more time on harder ones.

This makes answers clearer and more accurate, especially for math and coding, while still feeling fast in normal conversations. On top of that, OpenAI added detailed controls that let users change ChatGPT's tone and style.

Personality Presets

One big change is that ChatGPT now has different personality presets like Professional, Friendly, or Quirky, which affect how it talks across all chats. This directly responds to feedback that earlier versions felt too cold or robotic.

More Human Communication

By combining smarter reasoning with personality controls, ChatGPT feels more human and easier to use. It's less about raw intelligence now and more about how the AI communicates with people.

Google revealed Project Suncatcher, a long-term project that explores building AI infrastructure directly in space using solar-powered satellites. The idea is to use satellite constellations equipped with Google's TPUs and fast optical links to process data in orbit instead of on Earth.

This sounds extreme, but it makes sense when you realize how much energy the Sun produces and how much more efficient solar panels can be in space, where they get constant sunlight. Google is basically testing whether space could become the next place for massive data centers.

Proven Technology

Google has already tested key parts of this idea, including TPUs that survived intense radiation and satellite-to-satellite communication speeds fast enough for serious data transfer. Two prototype satellites are planned to launch by early 2027 in partnership with Planet to test how well everything works in orbit.

The Future of Computing

If launch costs continue to fall, space-based AI infrastructure could eventually cost about the same as Earth-based data centers. That would completely change how and where computing happens in the future.

Grammarly made a pretty unusual branding move by renaming its parent company to "Superhuman" after acquiring the email app with the same name. Instead of fully absorbing the Superhuman brand, Grammarly flipped the structure and made Grammarly itself a product under the Superhuman umbrella.

It's a confusing change at first, but also a bold one that shows the company wants to be seen as more than just a grammar checker.

Superhuman Go: The AI Assistant

Along with the rebrand, Grammarly launched an AI assistant called Superhuman Go, which connects to tools like Gmail, Google Drive, Calendar, and Jira. The idea is that the AI understands what you're working on and helps you write better while also automating small tasks, like logging tickets.

Competing with the Big Players

With earlier acquisitions like Coda and Superhuman, Grammarly is now building a full productivity suite instead of just a writing tool. This puts it in direct competition with platforms like Notion and Google Workspace.