Google spent Wednesday doing two things at once. In the afternoon it announced the strongest model it has ever built and told almost everyone they can't have it. After the market closed, it told investors it is taking ads inside AI Mode global for the first time. One hand closes the gate, the other opens the till.
That split is a useful way to read the whole day, because almost every story was about who gets access to AI and who gets billed for it. So instead of a rundown, here's how I'd sort it if I'm deciding what to do with my week.
Gemini 4 Carbon tops the charts and stays behind glass#
Gemini 4 Carbon arrives just two weeks after Argon, and Google is pitching it as a "significant leap" in coding, agentic work and especially cyberdefense. It takes first place on the Artificial Analysis Intelligence Index (v5.0), ahead of Claude Opus 5.5 and GPT-6 Astra, and leads SEC-bench Pro and ExploitGym, benchmarks that measure finding real bugs and turning them into working exploits.
The numbers Google chose to highlight:
0% success rate for prompt-injection attacks in the Grey Swan evaluation, against 0.7% for Argon, 1% for Opus 5.5 and 8.5% for GPT-6 Astra
Twice as likely as Argon to write code without vulnerabilities, and half as many "unsupported claims"
10x better at producing "high-quality" patches than non-Google models, according to the Chrome team
$2 per million input tokens and $10 per million output, an introductory price that doubles later
Google says internal teams have watched Carbon "run for days to persist through dead ends to tackle complex, multi-step engineering tasks." — via The Verge
Here's the catch. Right now only a small group of cybersecurity defense companies can use it. Google says the model's offensive potential calls for a "deliberative approach" through the voluntary US government testing framework, then a wider partner rollout. Ars Technica's headline calls it the frontier model "that you can't use", and it points out that Carbon beats Argon "by slim margins in every test", which might mean it's more of a post-training upgrade than a new model. If Argon's path is any guide, Gemini API developers could get it within weeks. Google also won't say when the discounted price ends.
My read: I'm not changing anything until it lands in the API. But the injection number is the one I'll test myself the day it does. Agents that read untrusted web pages and inboxes fail exactly there, and a model that really scores near zero would change how much scaffolding I build around tool calls. Benchmarks from the vendor are a claim, not a result.
What you can actually build on this week#
The more practical Google news is the Connected Apps Platform, now in public preview for US developers on a waitlist. You expose your service through a cloud-hosted MCP server, or through Android's AppFunctions APIs, then register, test and submit it in the Google Connected Console. You choose whether it shows up in Search, the Gemini app or both, and you get dashboards for traffic, latency and error rates. Gemini already has more than 50 connected apps, including Adobe, Experian and Peloton. They're coming to Search over the next few weeks.
That's the important part. If you already built an MCP server for Claude or ChatGPT, you now have a way into Google Search with the same protocol. This feels like early SEO for agents, and early is when the shelf space is cheap.
Nvidia shipped the other thing I'd try right away: Nemotron 3.5 Transcribe, a free, open-weight 30-billion-parameter speech model. It covers 40 languages, translates 31 to and from English, handles recordings up to an hour and tells up to eight speakers apart. It runs on a single chip, from a Jetson AGX Thor to a DGX Spark. It only works on recorded audio, not live streams, so voice agents still need something else. But for meeting notes or medical dictation that can't leave the building, it's worth trying against whatever hosted API you pay for today.
The meter is now the product#
Anthropic launched reporting tools that show managers what a team produced over a date range, what it cost and where workflows broke down, plus pooled group budgets for token spend. It's a dull feature with a sharp reason behind it. Jellyfish data cited in the same report shows spending on agentic coding tools up 10x since January, now about 2% of total engineering cost.
"Even if you're saying 2% is still small, that 10x growth is growth that most CFOs didn't plan for." — Nicholas Arcolano, Jellyfish
I'd turn these reports on even if nobody asked for them. The first time finance asks what the Claude Code bill buys, you want an answer that isn't a feeling.
Big Tech earnings explain why everyone suddenly cares about the meter. Microsoft booked $10 billion of revenue from OpenAI in a single quarter, almost 11% of its $92.3 billion total. Meanwhile OpenAI's unpaid balance to Microsoft rose to $7.2 billion. For the first time, Microsoft also put a dollar figure on Azure: $33.2 billion, up 48%.
Alphabet's numbers bring us back to that ads announcement. Google Cloud more than doubled to $30.7 billion, while Search growth slowed to 12%, down from 19% in Q1. Shares fell 8%. Hence ads in AI Mode, worldwide, plus multi-turn ads in the US. Google says sponsored listings in AI Mode already get about the same click-through rate as organic ones. If you're the one shipping connected apps into that surface, keep in mind that it's also an ad surface that is under pressure to make money.
What to watch from a distance#
Chrome's AI autofill can now pull data from Gmail, such as flight confirmations, order numbers and card numbers. It's opt-in and rolling out in the US. It's handy, and it's one more place where your inbox becomes context for an AI.
An xAI subsidiary is offering to buy homes in Southaven, Mississippi if the owners give up claims in a noise lawsuit against its data center and power plant. A realtor reportedly cited a target of 418 homes. The physical cost of compute is turning into a local political fight.
Here's my bet. By the time Carbon reaches the public API, the more lasting news from Wednesday will be the plumbing: an MCP route into Google Search, and an ad business that badly needs AI Mode to earn. The frontier model gets the headline. The distribution deal decides who gets paid.