- Google’s Carbon takes the benchmark crown (for now)
- Nate’s Notebook: Hire a dispatcher first
- Turn a prompt into a poster with Canva Craft
- An Anthropic manager bets on junior engineers
Google’s Carbon takes the benchmark crown (for now)
Image source: GoogleThe Rundown: Google just unveiled Gemini 4 Carbon, a new frontier model built for cybersecurity and long-running coding work that tops 15 of 17 benchmarks on Google’s charts, but isn’t yet public — with only select cyber defenders gaining access to start. The details: - Carbon’s 59 on Artificial Analysis’ “Intelligence Index” takes the top spot from Claude Opus 5.5 and improves 9 points over Google’s Argon from four weeks ago.
- Carbon’s widest wins over OpenAI’s Astra and Opus came on long-input reasoning (+24 points), hacking (+12), workflow automation (+11), and math (+8).
- Carbon’s $2/$10 intro API price (per million tokens in/out) is half of Opus 5.5’s and 1/5 of Astra’s, though it doubles to Opus’s rate after the initial discount.
- Google hasn’t given a date for wider access to Carbon, with Argon previously rolling out around two weeks after its debut to cyber defenders.
AI activity isn't the same as value 
The Rundown: 86% of CIOs say AI risk is growing faster than value — four times faster, on average. These Gartner IT Symposium/Xpo takeaways show CIOs how to compound value, not just activity, and govern AI before risk erodes advantage.Learn how to:- Balance AI value and risk
- Govern AI through an AI central bank
- Fill AI safety gaps with human oversight
Nate’s Notebook: Hire a dispatcher first
The Rundown: Each week, Rundown AI educator Nate Grahek breaks down what he’s seeing while teaching real teams to use AI.Nate: Most mornings, I brain dump my project list into my phone while I wait for coffee. Dot splits that one voice dump into separate chats running at once, each in its own context window. By the time I’m back at my desk, they’re gathering context, doing research, and drafting outlines.That’s the difference between the agents landing right now. Grok Bot and OpenClaw are monothreads: one worker running unattended. Even with careful rules and profiles, I’ve never been super impressed with that work. Dot is a delegator, handing discrete jobs to threads I can still coach.Unattended is the hard part. Think about hiring a new assistant: we all underestimate how hard it is for anyone, human or agent, to anticipate what matters to you. Agents will get there, but it takes time and a lot more context. Until then, I stay in the loop for chunking, quality, and coaching.Last week, I laid out my ladder: read, then write, then send. If you’re not ready for write access, that’s okay. Dot sorts the noise in my inbox and flags what needs me, with no access to write anything. Just letting it read and triage is a smart, useful, powerful first step, and it’s why I’m even more excited to finally get Anthropic’s Duet.No Dot yet? The brain dump works in a new ChatGPT voice chat, too.We’re currently in the messy middle. Triage and delegation are worth jumping into now; figure out the rest along the way.AI TRAINING
Turn a prompt into a poster with Canva CraftThe Rundown: In this guide, you’ll try Canva’s new image model on a workshop poster. You’ll check the headline, add event details you can edit, and make a square version without rebuilding the whole artwork.Step-by-step:- Open Dream Lab from Canva’s homepage and select Craft. Choose a portrait layout and check your AI allowance before generating
- Paste the guide’s brief. Ask for a clear headline and blank space for the event details. Then, check the headline very closely
- If the issue remains after one revision, generate the artwork without text and add the heading yourself
- Open the image in Canva and add the date, time, and venue as text boxes. Try Magic Layers only if you need to edit words or objects inside the generated image
- Make a square version and change the event date in both designs. Check the downloaded files for spelling, cropping, and the correct details
Just ask. Skip to done.
The Rundown: Eney by MacPaw is a proactive AI assistant that takes tedious work and life tasks off your list on Mac and iPhone. It spots what needs attention and handles tasks with your go-ahead.With Eney, you can:- Act on what’s on your screen with Eney Wand
- Delegate tasks, big and small
- Work seamlessly across your favorite apps
An Anthropic manager bets on junior engineers
Image source: Felix Rieseberg / J. M. W. TurnerThe Rundown: Anthropic engineering manager Felix Rieseberg wrote that he’s dropped the coding interview from his team’s hiring, saying Claude now outperforms people on the short coding tests it relied on — while also offering an optimistic counter-take that AI is letting junior engineers overtake veterans.The details: - Rieseberg, who has led engineering on Claude Code Desktop, says he has seen newer engineers “fly past” likely better coders, and that “hunger beats tenure.”
- His advice to juniors is to tie one bold idea to higher-ups’ goals and build it as a weekend prototype, since with AI “execution isn’t expensive anymore.”
- He used Claude Code’s desktop app as an example, with two newer employees prototyping it in one weekend before it went public just over a week later.
Trending AI Tools
QVAC SDK - runs any open-source model (Chat, Image, Voice Generation, and more), for free and privately. Try and star it on GitHub*
Nemotron 3.5 Transcribe - Nvidia’s AI that transcribes 40 languages
ChatGPT Dot - OpenAI’s always-on agent, now available via text message- GitHub Copilot Fusion - Copilot mode that auto-picks AI models for each task
Everything else in AI todayOpenAI is pulling hundreds of staffers into a “war room” for next week’s U.S. midterm elections, hunting for ChatGPT misuse after flagging Russian and Iranian influence campaigns on its models this month.Artificial Analysis released v5 of its “Intelligence Index” AI leaderboard, now relying more on private tests, with Chinese open models falling further behind in the rankings and OpenAI’s new Terra model outperforming despite its mid-tier level. New AI startup Particle released Particle-1, a DeepSeek V4.1 Flash variant trained to reason less on coding tasks, using ~45% fewer tokens with near-identical accuracy.Anthropic widened its government buying deal with a 40% Claude discount, saying civilian agency use is up ~20x in a year, after a judge blocked Trump’s order to drop it.Google opened a waitlisted preview letting U.S. developers build apps for Gemini and Search, with several partner apps also heading to Search within weeks.
Highlights: News, Guides & Events- Read our last AI newsletter: Google’s AI worlds can now talk back
- Read our last Tech newsletter: Apple’s big week kicks off with iPad mini
- Read our last Robotics newsletter: When 110 humanoids hit the dance floor
- Today’s AI tool guide: Turn a prompt into a poster with Canva Craft
Source: https://therundownai.beehiiv.com/p/goog ... ts-for-now