A features-and-tools week: Claude gets a browser, ChatGPT learns to move between temporary and permanent, Gemini transcription starts capturing intent instead of just words, and AI video crosses the faster-than-real-time line — all against a headlines backdrop where NVIDIA confirms its $12.9B Hugging Face buy and nearly posts a $100B quarter.
The line between AI assistant and AI operator just got a lot thinner.
No single frontier model dropped this week — instead, a dozen shipped features quietly changed how AI actually gets used. Claude got its own browser, ChatGPT Work got a computer in the cloud, Gemini's new transcription cleans up what you say into what you meant, and video generation now takes less time than watching the result. The quality-of-life upgrades are where AI moves from giving answers to opening the tab and getting the job done — and it's worth taking inventory.
If a task touches a website — filling out a form, working a web app, agentic browsing — a browser window now opens alongside your session for Claude to use. Nothing to install: it's built into the desktop app as Claude's own sandboxed browser, separate from your logins, with Claude-in-Chrome still there for those who prefer their own.
With Claude's browser, ChatGPT Work, and Grokbot all landing close together, the small updates add up to a meaningful inflection: AI that gives answers is becoming AI that opens the tab and gets the job done. a16z's Martin Casado draws the distinction sharply — ChatGPT Work is workflow-oriented 'fancy RPA,' while Grokbot, with its own browser and Slack chat, behaves more like a virtual coworker.
OpenAI's own examples for the feature — booking DMV appointments, setting up apartment utilities, comparing insurance policies, saving apartment listings — are strikingly consumer for a product called Work, an interesting wrinkle in the ongoing AI-for-life vs. AI-for-work divide. OpenAI's Romain Huet: 'ChatGPT Work now has its own computer in the cloud... the possibilities start to feel almost limitless.'
ChatGPT temporary chats can now be personalized with your existing memories, custom instructions, and plugins — and upgraded to saved history when they turn out to be worth keeping. Privacy-conscious users can default to temporary mode without giving up context.
A small update that punches way above its weight — power users juggling half a dozen inboxes have wanted this forever. Claire Vo recently called unconstrained connectors her favorite feature of the new Grok; as NLW puts it, 'I don't care who did it first as long as they all do it.'
This week: real-profile browsing, where the agent acts with your logins from a managed copy of your Chrome profile. That follows Bot Mode — named specialist bots with their own roles, models, memories, and skills that can talk to each other — plus a HUD overlay mode. Open source, any provider or model, a few dollars a month instead of two hundred.
A new set of plugins and skills lets Claude agents work the CRM directly — using Salesforce data as context, with agentic actions respecting existing governance guardrails via Salesforce's AI Force harness. Benioff: 'Here the UI is the AI... This is how every business will run.' It even got a new Matthew McConaughey ad.
Yes, 'our TAM is the US economy' lab announces a CRM integration — but Thomas Fanning says the deal made him rethink his priors: labs going through software vendors instead of around them signals that data, embedded workflows, and incumbency remain software's moats. The enterprise people NLW actually spoke to were extremely excited; one shelved plans to build a Salesforce alternative to see if this just does what they wanted.
The new speech-to-text model supports 85+ languages and offers verbatim mode or 'smart translate' mode, which strips filler, condenses rambling thoughts, and handles self-corrections — plus custom vocabulary for jargon, noisy-environment performance, up to three speakers, and an API. As NLW notes from constantly toggling Whisperflow's cleanup settings, that cleanup is the genuinely hard part — watch whether Google nails it.
Grok Voice Think Fast 2.0 now tops the Artificial Analysis speech-to-speech index, which measures whether voice agents can reason over what they hear and complete tasks with tools. At Starlink, it's diagnosing hardware issues, shipping replacements, and fulfilling over 3,000 orders a week via voice calls and chat.
The update lets you extend scenes, specify starting and ending frames, add video input references, upscale favorite takes to 4K, and test ideas quickly in 360p — massively more usability for generating video you can use for real things. It hit #1 in the text-to-video arena, though early testers are split on whether it clearly beats Seedance 2.5.
Fal's new model claims 3.49 seconds average latency versus 68 for Seedance 2.5 and 26.7 for Gemini Omni Flash — one tester generated a clip in 1.3 seconds for 20 cents. Ethan Mollick: 'A line in AI video was crossed' — it can now create reasonably high-quality video in less time than it takes to watch. Investor Steve Jang sees fully generated, massively multiplayer live visual worlds ahead. You can try it on venice.ai.
The Information broke the story late Wednesday after NVIDIA's earnings call. Hugging Face only has $150M in annualized revenue, so NVIDIA is technically paying around 80× — versus roughly 15× for SpaceX's $60B Cursor acquisition. But NVIDIA isn't buying revenue; it's buying the premier distribution channel for open models.
The pattern: a $6B Poolside licensing deal that brought in ~100 veteran AI researchers, a strategic Perplexity partnership for harnesses and app-layer exposure, and NeMo-Tron going from tech demo to real open-model contender in six months. It's also defense — as the biggest chip customers build their own silicon, open models downloaded, fine-tuned, and served on NVIDIA GPUs through CUDA are the counterweight.
Hugging Face is valuable partly because it's perceived as independent — its users include researchers, enterprises, cloud providers, and NVIDIA's own competitors. Eric Hartford calls the deal a loss for open source and 'an opportunity for a new standard bearer to arise.' Still, there's notably more optimism around this acquisition than usual — a sign of how much trust NVIDIA has built in the AI community.
Revenue growth accelerated 21 points over Q1, tantalizingly close to the $100B/quarter milestone only five other companies have hit. Guidance cools slightly — 89.5% for Q3 and 70% for fiscal 2027 on supply-chain snarls — but Amazon's new order for two million additional chips (total annual NVIDIA volume is believed under ten million) keeps demand solid for years. Stock up 4.8% after hours.
Free cash flow fell 50% to $21B while accounts receivable rose 50%, with days sales outstanding jumping from 45 to 60 — NVIDIA cites 'extended payment terms on large multi-quarter agreements with certain investment-grade customers.' Also disclosed: NVIDIA is officially back in China with H200 sales, though well short of what the administration permits.
Per the Wall Street Journal, the two-month-old program offering credit support in exchange for a cut of customer profits is on hold after employees flagged antitrust risk and sensitivities about how much control NVIDIA would exert over customer businesses. NVIDIA denies the reporting, saying the model 'is still in place and continues to evolve due to high demand.'
Agentforce is on track for $1.5B in revenue this year, up from a $1.2B forecast, and Salesforce is now just 5% from filling the entire SaaSpocalypse drawdown. Benioff's victory lap: seats grew across sales, service, and Slack, attrition near record lows, bookings more than doubled quarter over quarter, and agentic platform use surged sixfold. David Sacks: the 'SaaS is dead' narrative is 'getting shredded.'
Per The Information: Cognition hit $900M annualized revenue (more than 3× since January), Higgsfield tripled to $700M, Perplexity tripled to $750M, OpenEvidence doubled to $300M, and Manus — even after the messy Meta unwind — claims $400M, quadrupled from last year. DeepSeek's $70M this year is a 10× jump on its full 2025 numbers.
A drafted EO would have created an industry self-regulatory body modeled on FINRA — the idea Demis Hassabis proposed in July and Treasury Secretary Bessent publicly backed — requiring labs to share security notes, safety-test new models, and write industry rules. Reported opposition from David Sacks appears to have helped stall it. For now, no FINRA for AI.
— David Sacks, former White House AI czar, on FINRA for AI (All In podcast). Sacks argues that once the body exists, frontier labs will push for standards that apply equally to all models — standards open-model developers would find impossible to meet. He also notes FINRA's reputation among fintech startups: regulation that protects incumbents rather than promoting competition.
An open letter — signed by Visa, General Motors, and PwC alongside Google, Microsoft, and AWS — warns that AI-enabled cyberattacks will become far more widespread 'in the coming months,' putting hospitals, water treatment plants, and internet infrastructure at risk. It calls for defenders to get better AI tools and cross-industry coordination during a closing 'defender's window.' Ethan Mollick's warning: your organization isn't spending enough before open-weights frontier-class models and harnesses arrive.
— Sam Altman, on the industry cyber-defense open letter. Altman positioned security above competition in his response to the open letter: 'This is a critically important moment for cyber defense with AI. There is not much time to act... Only an urgent and intense collective response will work.'