Google held its 2026 I/O keynote at the Shoreline Amphitheatre this morning with a notable absence at the top of the deck: Gemini 3.5 Pro, the flagship model the developer audience had come to see, has slipped to June. Sundar Pichai’s stage-managed concession reportedly drew a groan from the room. That groan is the right frame for the rest of the day. The product Google shipped today is real and substantial. It is just not the product Google planned to ship.
What did ship: Gemini 3.5 Flash as the new default across Search and the Gemini app; Gemini Spark, a personal agent that takes actions inside Workspace; the first Android XR audio glasses from Samsung and Warby Parker; a redesigned, multimodal Search box billed as the biggest change to Search in over 25 years; and, quietly the most consequential announcement of the day, a **repricing of AI Ultra to 250 plan dropped to $200. The Pro slip is the headline. The pricing move is the news.
What 3.5 Flash actually does
Gemini 3.5 Flash is a Flash-tier model that posts Pro-tier benchmark scores. Google’s own published numbers have it beating Gemini 3.1 Pro on Terminal-Bench 2.1 (76.2%), MCP Atlas (83.6%), and GDPval-AA (1,656 Elo), with CharXiv Reasoning at 84.2%. It is roughly 4x faster on output tokens per second and lists at 9.00 per million output tokens, about 40% cheaper than 3.1 Pro. It is now the default in the Gemini app and in AI Mode in Search, which Google says now serves over a billion monthly users with queries roughly doubling every quarter.
The economic story this tells is more interesting than the benchmark story. A Flash model that outperforms the previous Pro on agentic and multimodal benchmarks, at 60% of the price and 4x the throughput, is Google quietly resetting the price-performance floor on the kind of work that drives most production traffic: coding, tool-calling, and search reformulation. Anyone running serious volume on 3.1 Pro should be running the migration math today.
The Pro slip is the cost of that story. Google appears to have decided that shipping Flash today and the Pro in June was better than holding both for a coordinated launch. From outside, that reads as either capacity discipline or training-run trouble; Pichai gave neither explanation. What it isn’t is a confidence move.
Gemini Spark is the actual product bet
If you strip away the demos, Gemini Spark is Google’s attempt to put Anthropic-style agent behavior inside its own consumer surface. Spark is described as a “personal agent” that takes actions across Gmail, Docs, and other Workspace apps on the user’s behalf: drafting, scheduling, summarizing, ordering things. MCP integration arrives this summer. It launches next week, in the US, only to the new AI Ultra tier.
The packaging matters. Spark is not an API. It is not a developer surface. It is a consumer agent gated behind a $100/month subscription that you talk to inside the same Gemini app you already use for chat. This is closer in shape to ChatGPT’s Operator than to Anthropic’s Claude Cowork. Google is selling the agent as an upgrade to the assistant, not as a new workspace.
Two things are worth noticing about that choice. First, by gating Spark behind Ultra rather than including it in Pro, Google is signalling that agentic action is a premium tier, not a baseline assistant capability. That contrasts with Anthropic’s recent SMB push, which bundles agentic workflows into a packaged tier sold on outcomes, not on model access. Second, the MCP integration arriving this summer is the actual extensibility story. Once Spark can call third-party MCP servers, the “personal agent” stops being limited to Google’s first-party surfaces.
The companion announcement is Android Halo, a top-of-screen indicator that shows what your agent is currently doing. It is the kind of small UX detail that gets dismissed and shouldn’t. Halo is the first serious attempt by a major OS vendor to give long-running agent work a persistent visual home outside of a chat window. Anyone who has tried to run a multi-step agent inside a tab and wondered whether it was still working will recognize the problem it solves.
The Search box redesign is bigger than it sounds
Google framed the new AI Search box as its biggest change to Search in over 25 years. That phrasing is doing some marketing work, but the underlying claim is fair. The redesigned box expands dynamically as you describe what you want, accepts text, images, files, video, and Chrome tabs as inputs in a single query, and is built on Gemini 3.5 Flash globally. Personal Intelligence (connecting AI Mode to Gmail, Photos, and soon Calendar) now reaches nearly 200 countries and 98 languages.
Read against the Gemini 3.1 Pro launch six months ago, the through-line is clear: Google has spent the last year repositioning Search from a keyword box to a query interface. The change you can see in today’s UI is the change they were already making in the model layer. Keyword search isn’t being deprecated, but it has been visibly demoted to a fallback.
Three other Search-adjacent features rolled out under the same umbrella. Information Agents can monitor blogs, news, social posts, and live finance and sports data 24/7 and surface what changed, coming this summer to AI Pro and Ultra. Mini Apps let users compose custom dashboards and trackers inside Search. Universal Cart is a shopping hub that gathers products across sites, finds deals, and flags incompatibilities; it lands in Search and the Gemini app this summer, then in YouTube and Gmail. These are all examples of Google using AI to move Search from a destination back toward an action surface. Every major AI assistant is trying to make that move; Google has more distribution to do it with.
The $100 Ultra tier is competitive pressure made visible
The pricing change is the least flashy announcement and the most informative. Google has restructured its AI plans:
- AI Ultra now starts at $100/month, with 5x the usage limits of AI Pro and access to Gemini Spark.
- The previous 200** with identical capabilities.
- The Gemini app is shifting from daily message caps to a compute-used model that allocates capacity by query complexity, refreshing every five hours up to a weekly maximum.
The price cut on the existing 100 introductory tier and the move to compute-based metering both read as responses to two pressures: a chat market where Anthropic and OpenAI are selling Pro-tier subscriptions at $20, and an enterprise market where Anthropic is selling Opus 4.7-class capability on per-token API pricing with no consumer subscription gate. The compute-based metering is the more interesting half. It is a tacit admission that “messages per day” is the wrong unit for a market in which one agent task can consume the equivalent of fifty chat replies.
For buyers, this changes the math. A small team that was deciding between Claude Pro and Gemini Advanced now has a credible $100 Ultra option with agentic features, image and video generation included, and substantially higher limits than Pro. The right call depends on workload, but the comparison just got tighter than it was last week.
The glasses, and what they’re actually for
Google spent the tail end of the keynote on the first Android XR intelligent eyewear: two form factors, one audio-only with cameras, microphones, and speakers, and one with an optional in-lens display for “private contextual information.” The hardware is engineered by Samsung and Qualcomm, with frames by Gentle Monster and Warby Parker; they work with both Android phones and iPhones; they ship this fall. No price was given.
The use cases shown were the ones you would expect: ask Gemini about what you’re looking at, get a restaurant review, translate a sign in real time, order food through DoorDash, send a message. None of them are new. What is new is that this is the first credible hardware bet from a major AI vendor on glasses as the next assistant surface. Meta has been here for two years, Apple has Vision Pro at the wrong price and weight, OpenAI has the Jony Ive partnership without shipping product, and Anthropic does not have hardware. Google has Samsung, Qualcomm, Warby Parker, and Gentle Monster as partners. That is a real consumer go-to-market.
The honest read on the glasses is that they’re a long bet. The category has not worked yet, and “Gemini in your glasses” doesn’t solve the social-acceptability problem that killed Google Glass. But the hardware partners and the Android XR platform together represent a serious attempt rather than a research demo, and the fall launch will be the first time we get to see whether the AI surface, finally, is good enough to justify wearing the thing.
What got the time, and what didn’t
Veo 3.1 and Lyria 3 got real but limited time on stage. Veo 3.1 is now free for any Google account with 10 video generations per month; Ultra accounts get 1,000. Lyria 3 and Lyria 3 Pro now generate full songs up to three minutes through the Flow Music platform’s “Chat with Producer” interface. These are substantive consumer upgrades and easy to undersell; they are also Google catching up to OpenAI’s Sora and Suno on creative generation rather than leading.
CodeMender, the DeepMind agent that finds and patches code vulnerabilities, got a fresh I/O segment despite first being announced in October. The substance hasn’t changed: it uses Gemini Deep Think, every patch is human-reviewed before submission, and it has upstreamed 72 fixes to open-source projects, including some with 4.5 million lines of code. The re-promotion suggests Google is positioning CodeMender as evidence that its agents do real, verifiable work, which is a useful argument against the broader skepticism that AI agents are demoware. Whether it ships as a product or stays a DeepMind research showcase is the part that still isn’t clear.
Workspace updates also got a slot. Gmail Live conversational search, Docs Live, Keep AI, the new Google Pics image app, and the Daily Brief personalized digest are mostly summer 2026 rollouts to AI Pro and Ultra. They are the kind of features that, if they work, become invisible inside three months of use.
What did not get the time: a Gemini 3.5 Pro demo, an updated Gemma release, anything substantive on Android 17, and any meaningful TPU or infrastructure update. The infrastructure absence is the one to note. Google’s pitch a year ago was the vertical stack from TPU to model to product. This keynote was almost entirely about the product layer.
The strategic question
I/O 2026 was not the model-launch keynote it was designed to be. The Pro slip forced a reshuffle, and what filled the time was a coherent push to make the Gemini app the dominant consumer AI surface: Spark as the agent, Ultra at $100 as the new high-end SKU, Search as the front door, and glasses as the next surface to colonize.
The strategic question is whether that adds up to a defensible position. Google has distribution other AI vendors do not have. It also has a model launch on its calendar that already slipped once, a pricing structure that just got cut to compete, and an agent product that ships with no third-party integrations until the summer. The next time Pichai has the room’s attention is the June Pro release. The bar for that demo just went up.