AI Developments

This Week in Google Gemini

Gemini is Google's family of AI models, the brain that now sits behind Search, Chrome, Workspace, and even the smart speaker on your kitchen counter. It's less one product than a layer Google is spreading across everything it makes, and this week showed just how far that spread has gone.

The headline move was the launch of Gemini 3.6 Flash and 3.5 Flash-Lite, two leaner, cheaper versions of the model built for speed rather than raw power — think of them as the difference between a sports car and a fuel-efficient commuter car doing the same daily drive. The new Flash model uses about 17% fewer output tokens than its predecessor, which in plain terms means it does the same job while burning less computing power, and it's noticeably better at coding without going on the kind of "let me try that again" loops that used to trip up earlier versions. Google also teased that a bigger Gemini 4 is on the way, without saying much more than that. Meanwhile, Gemini in Chrome got a redesigned side panel that lets you multitask with the AI while browsing instead of switching tabs, and Google struck its first-ever AI news deal with a publisher, folding Associated Press journalism directly into Gemini's answers. Even Gemini for Home got an upgrade, with its memory of a conversation now stretching to 15 minutes across Nest speakers and displays.

Taken together, it's a week about Google making Gemini faster, cheaper, and more woven into daily life — your browser, your smart speaker, your search results — while quietly building anticipation for whatever Gemini 4 turns out to be.

This Week in Meta Llama

Llama is Meta's line of AI models, historically famous for being free and open for anyone to download and build on, which is a big reason it powers Meta AI across WhatsApp, Instagram, and Messenger for billions of people. That open, give-it-away spirit is exactly what's being tested right now.

There's no fresh Llama-branded release to report this week — the next major model, sometimes called Llama 4.5 or codenamed "Avocado," is reportedly still stuck in development, with Meta's Behemoth version that was expected back in June still not out the door. What is happening is a bigger strategic shift happening around Llama rather than to it. Meta has been quietly pivoting away from its open-source roots toward closed, paid models built by its new Superintelligence Labs team under Chief AI Officer Alexandr Wang. The clearest evidence is Muse Spark 1.1, a proprietary coding-and-reasoning model that Meta started charging developers to use for the first time ever, and which reportedly beat Google's latest Gemini release on some coding and reasoning benchmarks. Meta has also floated a "Llama API" developer platform meant to blend the openness people loved about Llama with the tighter control of closed models like Muse Spark.

Quiet week for Llama itself — no major model releases in the past seven days. The last notable move was Meta launching Muse Spark 1.1 on July 9, its first paid AI model and a sign the company may be stepping back from Llama's free, open-source tradition.

This Week in Microsoft Copilot

Microsoft Copilot is the AI assistant Microsoft has been stitching into basically everything — Word, Excel, PowerPoint, Outlook, Teams, even Windows itself. Think of it less like one chatbot and more like a helper that follows you from app to app, and lately Microsoft has been spending most of its energy on letting companies build their own custom helpers on top of it.

The biggest shift this month is that Copilot's "Agent Builder" now lets employees submit the mini AI agents they've built to a company-wide Agent Store, in a section literally called "Built by your org." An admin has to review and approve them first, kind of like an app store gatekeeper, but once approved, anyone at the company can find and use them. Alongside that, Microsoft rolled out governed agent publishing and tenant-wide prompt galleries, meaning a whole company can now share a library of pre-written prompts and vetted agents instead of everyone reinventing the wheel. Copilot also picked up broader access to MCP agents — a kind of universal plug for connecting AI to outside tools — across Word, Excel, PowerPoint, Outlook, and Catalyst, plus tighter controls over which connectors and permissions those agents can actually touch.

On top of that, Copilot Chat now serves up richer Bing web-answer cards directly in conversation, there's a new customizable landing page for an Employee Self-Service agent, fresh admin controls for AI video generation, and easier one-click sharing of agents straight into Teams. Microsoft even renamed its Copilot Studio certification track to "Microsoft 365 Copilot specialization," swapping out an old admin exam for new agent-building credentials, which tells you where the company thinks the real skill gap is right now.

Put together, this is Microsoft betting hard that the future isn't one giant AI, but thousands of small, purpose-built agents that regular employees build and share like office memos. If that bet pays off, using AI at work starts looking less like typing into a chatbot and more like browsing an internal app store for exactly the helper you need.

This Week in Artlist

Quiet week for Artlist — no major product announcements in the past seven days. The last notable moves were the April launch of Artlist Studio, its AI-powered video production platform, and the company crossing $300M in annual revenue before announcing in June that it was cutting about 200 jobs (roughly 40% of its staff) as part of a shift toward what it called an "AI-native operating model."

This Week in SAP Joule

If you've never heard of SAP, think of it as the software running the back office of a huge chunk of the world's biggest companies — payroll, supply chains, expense reports, the works. Joule is SAP's AI assistant sitting on top of all that, letting employees just type or talk instead of hunting through menus. And this week, Joule quietly got a lot more hands-on with the boring-but-essential stuff that actually runs a business.

The biggest theme was Joule showing up in places it never used to be. It can now handle Concur Expense reports end to end — creating, editing, submitting, even recalling or deleting them — just by chatting with it in plain language. Booking a work trip got the same treatment: Joule can now search flights, compare fares, pick a hotel, add a rental car, and walk you through checkout, all through conversation instead of a maze of dropdown menus. On the more technical side, developers working in SAP's cloud infrastructure got a Joule sidekick too, one that can inspect Kubernetes clusters, pull pod logs, and search documentation, which is the kind of troubleshooting that used to eat up an engineer's whole afternoon. SAP also pointed to real numbers behind these features: developers using Joule to automate coding tasks saw a 20% productivity bump, and one retailer, LC Waikiki, used a Joule-built HR agent to cut process times by 40 to 60%.

None of this is flashy the way a new chatbot demo is, but it's arguably more telling. SAP isn't trying to make Joule impressive in a keynote — it's trying to make it disappear into the plumbing of everyday work, one expense report and one Kubernetes dashboard at a time. For anyone whose job involves SAP software, that likely means fewer clicks and less form-filling in the months ahead.

This Week in Einstein AI

Salesforce Einstein started life as the "smart suggestions" layer inside Salesforce's CRM — the thing that predicted which lead might close or which email to send next. These days, Einstein has largely been folded into Agentforce, Salesforce's bigger push toward AI agents that don't just suggest things but actually go do them. This week's news is really about how far that shift has come.

Salesforce's Spring '26 release pushed Agentforce deeper into nearly every corner of the platform, adding things like a beta version of Setup powered by Agentforce, a fully available "Agentforce 360" bundle, a redesigned Sales Workspace, two-way messaging, and even a built-in ChatGPT integration. Underneath all of it, Einstein's prediction engine is now paired with genuine agent tools — an Agent Builder for assembling custom agents, orchestration so multiple agents can work together on one task, and human-in-the-loop checkpoints so a person can step in before anything risky happens. One small but telling update: the Prospecting Agent now lets admins tune how an agent behaves just by typing plain-English instructions, with a live preview showing the effect immediately, instead of digging through a builder interface. Salesforce also isn't married to its own AI anymore — agents can run on OpenAI's models by default, switch to Claude for regulated industries like healthcare or finance, or use Google's Gemini, depending on the job.

The headline number worth remembering is that Salesforce says over 12,000 companies are now live on Agentforce, with many reporting 30 to 50% less time spent on manual tasks. That's the real story here: Einstein isn't just predicting your next move anymore, it's often the one making it.

This Week in ChatGPT

ChatGPT is OpenAI's chatbot, the one that kicked off this whole AI boom, and these days it does a lot more than answer trivia questions — it drafts emails, plans trips, and increasingly acts like a work assistant. This week OpenAI leaned hard into that "assistant" identity, rolling out a cluster of updates aimed at making ChatGPT feel less like a chat window and more like a tool you actually live inside.

The biggest shift was on the desktop app, which got a redesign splitting things into a clearer "Chat" side and a "Work" side, with your Projects now showing up right in the app and Work conversations syncing across web, mobile, and desktop — so you can start a thread on your laptop at the office and pick it back up on your phone on the train home. Alongside that, OpenAI launched a ChatGPT for Small Businesses program, complete with training sessions and in-person "AI academies," pairing it with wider access to GPT-5.6 and the newer Work agent. It's a clear signal that OpenAI wants ChatGPT to be the default tool small business owners reach for, not just something they play with. On the personalization side, custom instructions jumped from 1,500 to 5,000 characters for Plus, Enterprise, Business, and Education users, which sounds like a small technical tweak but actually means you can give ChatGPT a much richer, more detailed sense of who you are and how you like things done before it writes a single word.

Put together, the week's news is less about a flashy new model and more about ChatGPT settling into daily life — work, small business, personal routines — with the plumbing to match. If you use ChatGPT regularly, the practical upshot is a smoother experience jumping between devices and a chance to fine-tune it to actually sound like it knows you.

This Week in Mistral AI

Mistral AI is the French AI lab that's built a reputation as Europe's answer to OpenAI and Google — known for efficient, often open-weight models and a chat app called Le Chat. This week's biggest story wasn't really about a model at all, though: it was about money and muscle.

On July 21, Microsoft and Mistral announced a major expansion of their partnership, a deal reported to be worth billions, giving Microsoft access to Mistral's European computing infrastructure while helping regulated industries adopt AI they can actually control and keep on European soil. That plugs into a bigger buildout story for Mistral, which is also standing up a new 10-megawatt data center facility near Paris, due to open in the third quarter, and has been working with heavyweight industrial partners like Airbus, BMW, and ASML on AI tools for engineering and manufacturing. On the product side, Mistral's agentic assistant, Vibe, keeps getting more capable — it's now positioned as a single agent that can comb through your inbox and calendar, do deep research, draft documents, and even take a coding task all the way from request to a shipped, reviewed pull request. There was also a quieter but useful update: a bugfix release for Mistral's underlying tokenizer tooling that cleans up some technical rough edges around audio processing and text decoding.

The throughline here is Mistral positioning itself as the "trustworthy, sovereign" AI option for European businesses and governments, backed by real infrastructure investment rather than just marketing. For everyday users, the more interesting thread to watch is Vibe, which is quietly becoming a genuine competitor to the do-everything agents coming out of the US labs.

This Week in Claude Code

Claude Code is Anthropic's command-line tool that lets developers hand off actual coding work to Claude instead of typing every line themselves — think of it as a very capable junior engineer who lives in your terminal. This past week, Anthropic spent most of its energy making that junior engineer faster, calmer, and easier to trust.

The headline fix tackles something that was quietly annoying power users: on long coding sessions, the tool used to get slower and slower the more you talked to it, because behind the scenes it was reprocessing the entire conversation history every single turn — a bit like re-reading an entire group chat from the beginning every time someone sends a new message. Anthropic squashed that slowdown, so marathon sessions now stay snappy instead of grinding to a crawl. Alongside that, Claude Code picked up new sandbox controls that let teams skip full filesystem isolation while still keeping a lid on network access, giving companies more control over exactly how much freedom to give their AI coding assistant.

On the workflow side, Claude Code now runs code reviews in the background automatically, and screen-reader users got a much richer, more detailed narration of what the tool is doing — a real accessibility upgrade rather than an afterthought. Anthropic also beefed up "auto mode" and trust handling so the assistant is more careful about what it's allowed to touch on its own, and cleaned up a long list of smaller bugs around sessions, permissions, worktrees, and even how text renders inside VS Code. For teams managing lots of developers, the admin console also got two new tabs that track usage and value, so a manager can now see things like active developers, session counts, and which commands get used most, updated daily.

None of this is flashy, but it's the kind of unglamorous polish that decides whether a tool becomes part of someone's daily habit or gets abandoned after a rough first week — and this week Anthropic clearly bet on habit.

This Week in Grok

Grok is xAI's chatbot and AI model family, built by Elon Musk's company and baked directly into X — it's known for a slightly edgier, more "say what it thinks" personality than some of its rivals. This week Grok pushed forward on two very different fronts at once: making itself more useful for everyday automation, and opening up its coding tools to anyone who wants to peek under the hood.

The biggest practical addition is Automations — you can now tell Grok to run a task on a schedule, or automatically whenever a matching email shows up, and it'll report back through email or a phone notification. It's basically giving Grok the ability to work while you're not even looking at it, and it's available right now on grok.com and the mobile apps, with the email-trigger option reserved for SuperGrok subscribers. On the developer side, xAI open-sourced Grok Build, its coding agent and terminal interface, meaning programmers can now inspect, modify, and run the whole thing themselves instead of trusting it as a black box — a notably generous move for a company that's usually pretty guarded about its tech. That release also came with smaller but welcome polish, like clearer session details and better handling of rate limits and queued messages.

Meanwhile, on the model front, xAI shipped Grok 4.5, a massive 1.5-trillion-parameter model tuned especially for coding, and Musk declared that Grok Imagine, the image and video generation half of Grok, has officially reached completion, showing off some of its artistic output on X. The much-hyped Grok 5, though, still doesn't have a confirmed release date, with most signs now pointing to sometime in the third quarter of 2026 or later.

Put together, it's a week where Grok looked less like a single chatbot and more like a growing toolkit — one piece for automating your inbox, one for coding, one for generating art — with the biggest model upgrade still sitting just out of reach.

This Week in Claude Cowork

Cowork is the part of Claude that acts less like a chatbot and more like a coworker — it can open files, manage tasks, and work through multi-step jobs on your computer instead of just answering questions in a text box. Up until now, it lived only inside the Claude desktop app, which meant you had to be sitting at that one machine to use it.

That changed this week. Anthropic announced Cowork is breaking out of the desktop app and heading to the web at claude.ai and to mobile on iOS and Android, starting in beta with Max plan subscribers. The bigger deal isn't just "now on your phone" — it's how it works under the hood. Sessions now run remotely on Anthropic's own servers instead of your laptop, so your files and progress are tied to your account, not your device. Close your laptop mid-task, and Cowork just keeps working. Scheduled jobs, like this very digest, can now fire off even if every one of your devices is powered down. On top of that, chat and Cowork got merged into one shared home screen, so you're not toggling between separate apps for separate modes anymore, and new Microsoft 365 write tools let Claude draft and send emails, manage your calendar, and edit OneDrive or SharePoint files directly.

Think of it like the difference between a work laptop that only runs your projects when it's open, versus a project that keeps humming along on a server somewhere no matter where you are. For anyone juggling tasks across a phone and a desktop, that's the actual unlock here: less "let me get back to my computer," more "it's already done by the time I check."

This Week in Eleven Labs

ElevenLabs is the company known for eerily realistic AI voices — the kind that power audiobooks, voice assistants, and dubbed videos that don't sound robotic. Lately it's been stretching beyond just voice generation into full conversational AI agents that businesses can deploy for calls and chats, and this week's updates lean hard into making those agents easier to actually manage at scale.

The headline change is conversation tags: a proper system for labeling, filtering, and organizing the (potentially huge) pile of conversations an AI agent has with customers, so a business can sort through what happened without manually scrolling through transcripts. Alongside that, ElevenLabs added filters to hide conversations by status, so teams can quickly zero in on the ones still "in progress" or the ones that failed and need a human look. The platform also plugged in fresh underlying language models, including newer Claude and GPT options, giving developers more choice in what "brain" powers their voice agent. Smaller but useful additions rounded things out too, like better real-time transcription, auto-translated transcripts, sentiment analysis on conversations, and longer timeouts for tools connecting through MCP, the increasingly common standard for linking AI to outside data and services.

None of this is flashy on its own, but together it reads as a company shifting from "cool voice demo" to "enterprise-ready tool." If you're a business running voice or chat agents at any real volume, this is the unglamorous plumbing work that makes that actually manageable day to day.

This Week in Anthropic Claude

Claude is Anthropic's AI assistant — the one that writes, codes, researches, and increasingly runs whole workflows on your behalf. This week it leaned hard into two very different audiences: teachers and IT admins.

Claude for Teachers launched, giving verified K-12 educators in the US free access to premium Claude tools, ready-made teaching skills, and curriculum connections tied to standards in all 50 states, plus training to help teachers actually get comfortable using AI in the classroom. Think of it as Anthropic handing teachers a customized toolkit instead of just a generic chatbot. On the other end of the spectrum, Claude Code — the coding-focused version of Claude — picked up a big stability and safety update: tighter permission checks, safer handling of terminal commands, cleaner cleanup of background sessions, and a new "EndConversation" tool with progress heartbeats so long-running coding tasks don't just vanish into a black box. Meanwhile, Anthropic beefed up the business side too, adding user-management tools to the Enterprise Admin API (so IT teams can manage roles and groups without digging through settings by hand) alongside richer analytics, spend alerts, and model-level permissions for companies running Claude at scale. There's also a new personal touch: a "Reflect" dashboard and monthly recap showing you your most active day, peak hour, and what you've actually been using Claude for.

Put together, it's Anthropic building out both ends of the ladder — making Claude more approachable for a teacher who just wants help grading essays, while making it more controllable and auditable for a company running hundreds of automated coding agents. That's usually a sign a tool is maturing: less "cool demo," more "everyday infrastructure."

This Week in Perplexity AI

Perplexity started as an AI search engine that actually shows its sources, but it's been racing to become something bigger: a full-on AI agent that does tasks for you, not just answers questions. This week's news doubles down on that shift, and one announcement in particular sounds like something out of a sci-fi movie.

Perplexity unveiled "Personal Computer" — an always-on AI that lives on a dedicated Mac mini in your home or office and works around the clock as a kind of digital proxy, watching for triggers and quietly handling tasks even while you're asleep. Powering it is a new memory system called "Brain," which builds a private map of everything it's learned across your files, apps, and past decisions, then refreshes itself overnight so each new task starts already knowing what worked before and what didn't; in testing, that memory boosted answer accuracy by 25% and recall by 16% while actually cutting costs. On the business side, Perplexity's Comet browser — the one with a built-in AI assistant that can research, summarize, and even fill out forms for you — is now available for enterprise customers, and the developer-facing Agent API added support for more outside models, including GPT-5.4 and Gemini 3.1 Pro, plus a new finance-focused search tool.

The throughline here is Perplexity betting that people don't just want faster answers — they want an assistant that remembers context and keeps working when they're not looking. That's a meaningfully different pitch than "search engine with citations," and worth watching as more of these always-on agents show up on regular people's desks.

This Week in Google Gemini

Gemini is Google's flagship AI, baked into everything from Search to Docs to your Android phone. It's less a single app and more a layer of smarts spreading across the whole Google universe, and this week that spread got noticeably wider.

The biggest buzz is around Gemini 3.5 Pro, which insiders say was targeting a July 17 launch after Google DeepMind reportedly scrapped an earlier build when engineers found it stumbling on things like recursive tool-calling (basically, an AI using tools to call other tools) and generating clean SVG graphics. Google hasn't confirmed a release date, a rumored 2-million-token context window, or pricing, so treat it as "coming soon" rather than "here now." Meanwhile the stuff that is confirmed is quietly useful: Gemini's video tool inside Google Vids can now do real editing work, not just generate clips, letting you ask it to fix physics glitches, clean up on-screen text, change the color grading, or strip out background noise with a simple text prompt. Google Sheets and Docs also expanded their Gemini-powered writing and formula help to eleven more languages, including Mandarin, Dutch, Hebrew, and Polish, so this isn't just an English-speaking upgrade.

There's a smaller but sweet story too: Gemini launched a pilot called ATL Saathi, a 24/7 planning assistant for teachers running "Tinkering Labs" in 100 Indian schools, offering lesson ideas and support in multiple languages. It's a reminder that alongside the flashy model races, Gemini is also being pointed at very grounded, everyday problems. If 3.5 Pro does land soon, expect a genuinely bigger jump than usual, since Google seems to be taking its time to get the fundamentals right before shipping.

This Week in Meta Llama

Llama used to be Meta's calling card: free, open-weight AI models that anyone could download and build on, in contrast to the locked-up systems from OpenAI or Google. That era looks like it's ending, and this week made the shift even clearer.

Back in April, Meta's new Superintelligence Labs division, led by AI chief Alexandr Wang, quietly replaced Llama with a new model family called Muse Spark, and this week they pushed out an update: Muse Spark 1.1, which Meta is calling its strongest model yet for agentic and coding work (meaning it's built to take actions and write code, not just chat). It's rolling out through a developer portal in public preview, so outside programmers can start testing it and hooking it into their own tools. The catch is that unlike Llama, Muse Spark is proprietary, meaning Meta controls access instead of handing over the blueprint for free. Wang has said the company still intends to open-source a variant of Muse Spark down the road, but for now the flagship version is closed. This all lines up with reporting from earlier in July that Meta is also pushing into the AI coding assistant market, chasing the ground that Anthropic and OpenAI have staked out there.

For anyone who built their projects on open Llama models, this is worth watching closely. It doesn't mean Llama vanishes overnight, but Meta's center of gravity has clearly moved toward a more locked-down, business-first approach, and the promised "open" variant of Muse Spark hasn't shown up yet.