AI the News That's Fit to Prompt

Google used a single I/O keynote to declare the ten-blue-links era over, ship a new flagship model (Gemini 3.5 Flash) that costs three times its predecessor, wire a 24/7 agent into Gmail, point a world model at Street View's 280 billion images, and re-tier its developer plans, all on the same day Anthropic announced it had hired Andrej Karpathy onto its pre-training team and bolted Claude into every desk at KPMG. Two labs, two strategies, one underlying bet: the next round of AI competition is not about chatbots, it is about who owns the long-running agents that sit in front of your work, your search, and your codebase.

Search and the End of the Ten Blue Links

• Google declares the link era over. AI Overviews already serve 2.5B users a month and conversational search has crossed 1B. Liz Reid says new "information agents" will "map out a monitoring plan for you, including the tools and the data it needs to access," extending the 2003 Alerts product into a continuously running background process. Generative UI will assemble custom interactive answers, and users can spin up "mini apps" via natural language through Antigravity.
• Publishers brace for another step down. Sundar Pichai framed the goal as bringing "frontier models to as many people as possible" at lower cost; the implicit cost is to anyone whose business model still depends on outbound clicks from Google.

The I/O Agent Stack

• Gemini 3.5 Flash is the engine. Google's DeepMind chief technologist said it "outperforms our latest frontier model, 3.1 Pro, on nearly all the benchmarks," runs "4x faster than other frontier models" (12x in an optimized variant), and is paired with a forthcoming 3.5 Pro that plans while Flash executes. Demos showed agents spawning subagents to build an operating system inside Antigravity.
• Spark is the consumer agent. Built on 3.5 Flash and Antigravity's harness, Spark runs on dedicated Google Cloud VMs, accepts work via a dedicated Gmail address, and is tracked through Android's Halo system. Josh Woodward: it "can pull all the facts from your emails, your docs, your sheets, and slides and write the draft for you." Spark goes to AI Ultra subscribers next week.
• Antigravity 2.0 is the developer surface. A new desktop app, CLI, voice control, and SDK for custom subagents, plus a pricing reset: a new $100 Ultra plan with 5x higher limits and the old $250 tier cut to $200 with 20x higher limits. Notably, Gemini 3.5 Flash was "co-developed using Antigravity."
• Omni is the media engine. Gemini Omni Flash spins images, audio, and text into 10-second video with multimodal reasoning (a single text prompt produced a narrated claymation protein-folding short). SynthID watermarks every output; an avatar onboarding flow gates likeness use.
• Genie sees the real world. Project Genie now ingests 280 billion Street View images from 110 countries to simulate real streets with weather and rare-event controls. Waymo already uses Genie 3 to train on tornado-grade events, though demos still betray weak physics (a simulated woman walks through cacti).

Anthropic's Quieter Counter

• Karpathy joins pre-training. The OpenAI co-founder and ex-Tesla AI lead is now on Nick Joseph's team and will stand up a group using Claude to accelerate pre-training research itself. The framing matters: Anthropic is betting AI-assisted research beats raw GPU count.
• KPMG goes all-in on Claude. 276,000 employees across 138 countries get Claude access, Digital Gateway gets it baked in, and Anthropic is named preferred partner for PE portfolio rollouts. Vice chair Rema Serafi: building a client-facing tax agent went from "weeks" to "minutes." KPMG's Blaze tool uses Claude Code for legacy modernization.
• Widening the conversation. Anthropic is convening 15+ wisdom-tradition groups to inform Claude's "moral formation," and reports that simply letting Claude recall ethical commitments mid-task produced "markedly lower rates of misaligned behavior" in evaluations, a small but striking interpretability result tucked inside a policy post.

Pricing, Politics, and Pauses

• Flash is no longer cheap. Simon Willison flags that Gemini 3.5 Flash is "3x the price of 3 Flash Preview and 6x the price of 3.1 Flash-Lite," at $1.50/M input and $9/M output, brushing 3.1 Pro pricing. One Artificial Analysis benchmark run cost $1,551.60. All three frontier labs are now "probing the price tolerance of their API customers."
• Stratechery doubles up on context. Ben Thompson's "Shifting Alliances" connects the answer-vs-agentic compute split to Anthropic's xAI deal, calling it evidence "markets actually work quite well," and cautions that "upper hand" narratives in US-China AI tend to be overblown.
• And a personal day. Thompson published no Daily Update on the loudest news day of the quarter, an inadvertent reminder that not every shift needs an instant take.

The Throughline

Strip the brand names off today's announcements and the architecture is the same shape on both sides. Google ships Search, Spark, Antigravity, Omni, and Genie as five surfaces, but the runtime under all of them is one model family (3.5 Flash now, 3.5 Pro soon) and one execution pattern, Pro plans, Flash acts, agents run for hours on cloud VMs you do not have to keep awake. Anthropic responds by buying the people who can compress that training loop further (Karpathy) and by colonizing the work surface from the other direction, dropping Claude into 276,000 KPMG seats and naming itself the default for PE portfolio rollouts. Different go-to-markets, identical thesis: the unit of value is no longer a model response, it is a persistent agent.

That thesis explains the pricing news as much as the product news. Willison's note that 3.5 Flash costs 3x what 3 Flash did, and that one benchmark sweep ran past $1,500, is the inevitable other side of the same coin. If you are billing for a chat turn, you compete on price per token. If you are billing for an agent that runs for hours and touches your inbox, your code, your slides, you are billing for a job, and the labs have decided customers will tolerate it. The cheap-tokens era is ending in plain sight, and the I/O keynote is the price tag for the agents Google just gave you.

The KPMG and Karpathy stories are the same story told in headcount. KPMG is interesting not because a large firm bought enterprise AI seats, but because the framing has flipped: Anthropic is now the platform that 276,000 professionals build agents on top of, and KPMG's Blaze tool is in essence a managed Claude Code service for legacy modernization. Karpathy moving to a pre-training team that will explicitly use Claude to accelerate Claude is the research-side mirror of that flip. The thing being scaled is no longer just the model, it is the human-plus-model loop, and both labs are now optimizing the loop.

Anthropic's "widening the conversation" post is the soft footnote that may matter more than it looks. The empirical claim, that giving Claude a tool to recall ethical commitments mid-task lowered misaligned behavior, lands in an industry that just decided to give agents hours of autonomous runtime over your data. The race for surfaces and the race for agent self-restraint are now happening in parallel, and the second one is being underfunded by everyone except, on the evidence of today's posts, Anthropic.

The Bigger Picture

I/O 2026 is the moment Google stops hedging. For two years the company described AI Overviews as an enhancement to Search; today it described Search as something that no longer exists in the form anyone learned to use. That is a generational reset of the most lucrative product in consumer software, and Google is doing it on a model (3.5 Flash) it shipped to "billions" of free users the same morning. The bet is that owning the agent layer end-to-end (model, runtime, surface, distribution) is worth disrupting your own cash cow. The corollary is that the open web, which has spent twenty-five years tuned to Google's link economy, has just had its primary referrer announce that referral is no longer the default behavior.

Step back further and the I/O stack plus Anthropic's plays sketch the shape of the next eighteen months. Compute is bifurcating into "answer inference" and "agentic inference," exactly as Thompson has been arguing, and the price curve confirms it: short, cheap turns get cheaper through commoditization while long, expensive agent runs get more expensive because customers will pay for the work, not the tokens. Enterprises will not stitch this together themselves. They will buy it as a bundle, which is why a KPMG-Anthropic alliance lands the same week as Google's everything-is-an-agent keynote. The platform layer is consolidating around two or three players, and the consulting layer is racing to be the integration partner before that window closes.

For builders, the practical message is harsh and clear. Anything you ship that assumes a search-driven distribution model, a per-token cost basis, or a single-turn UX is now on borrowed time. The defensible designs are ones that treat the agent as the primary user (Economist-style structured content, MCP-friendly APIs) and the human as the supervisor (Spark's Halo trail, KPMG's judgment-plus-AI framing). The labs have decided where the next contest lives. The rest of the industry has roughly one product cycle to decide whether to build for that world or be optimized around by it.

What to Watch

• Whether Search traffic actually falls off a cliff. AI Overviews and conversational search have been ramping for a year, but generative UI and information agents are a step change. Publisher analytics over the next 60 days will tell you how steep.
• How Anthropic answers the I/O stack. Karpathy is a signal, not a product. The thing to watch is whether the next Claude release uses pre-training research that was itself accelerated by Claude, the first visible iteration of the loop Anthropic is now staffing for.
• The price tolerance curve. If a 3x Flash hike sticks, expect OpenAI and Anthropic to follow within a quarter. If API customers revolt, expect a "Flash-Lite" reissue. Either way, the cheap-tokens era is over and the agent-runtime era is being priced in.