Google’s annual I/O kicked off on 19 May with the most ambitious AI keynote it has ever delivered, packing more than one hundred announcements into three days and re-centering the entire company around agents. Sundar Pichai opened by declaring this an agentic era, and the product reveals immediately backed up the framing.
At the model layer, Google unveiled Gemini 3.5 Flash, which it claims runs roughly four times faster than other frontier models while beating its own Gemini 3.1 Pro on Terminal-Bench 2.1 (76.2%), GDPval-AA (1656 Elo) and the new MCP Atlas (83.6%) agentic benchmark. Gemini 3.5 Pro is in private testing and will roll out next month. Above the model line, Google introduced Gemini Omni, a native any-to-any multimodal system that DeepMind CEO Demis Hassabis described as a world model rather than a video generator.
For developers, the company shipped Antigravity 2.0, a re-imagined desktop app, a new CLI and an SDK that let teams orchestrate multiple sub-agents and schedule them to run in the background. A single Gemini API call now provisions a managed remote Linux sandbox with browsing, code execution and tool calling. On the enterprise side, the Gemini Enterprise Agent Platform now processes 16 billion tokens per minute, and Google announced a $750M partner fund to accelerate agent deployment.
Google has finally matched the breadth of OpenAI and Anthropic’s agent stacks. For two years Google looked like it was always reacting. This week it shipped a frontier-class fast model (3.5 Flash), a world-model media stack (Omni), a redesigned developer environment (Antigravity 2.0), a managed agent runtime, and an enterprise platform doing 16 billion tokens a minute. Add the Magic Pointer demos from last week and the Googlebook hardware coming in autumn 2026, and the consumer surface is also there.
The strategic read is that Google now has the only end-to-end stack across silicon (TPU), models (Gemini), agents (Antigravity, Managed Agents), distribution (Search, Workspace, Android, Chrome, soon Googlebook) and enterprise (Gemini Enterprise plus the $750M partner fund). OpenAI has the brand and DeployCo. Anthropic has the enterprise trust and the Big Four. Google now has everything else.
The bottom line: watch take-up of Antigravity 2.0 and the new Managed Agents APIs in the next sixty days. That is the early signal of whether the demo bench strength translates into the paid pipeline OpenAI and Anthropic already enjoy.
Agentic AI consolidated as enterprise infrastructure this week. Google’s I/O blitz collapsed model, agent, IDE and enterprise platform into a single stack. Anthropic crossed into nine-hundred-billion-dollar territory at the same moment KPMG put Claude on every desk. Hark pulled seven hundred million dollars at a six billion dollar valuation for a brand-new universal AI hardware bet. Spotify, Microsoft and SpaceX all made distribution-led AI moves on the same day.
Agents are no longer a feature. They are the operating layer that buyers, builders and capital are now racing to own. The story has shifted from who has the best model to who owns the surface, the workflow and the partnership it ships through.
Omni’s pitch is that video generation has stopped being a separate model class and has become an output of a general reasoning system. By bundling generation, world reasoning and reliable provenance into one model, Google is positioning Omni against ByteDance’s Seedance, Kuaishou’s Kling and OpenAI’s now-shuttered Sora consumer product. It is also the model that finally gives Google a unified creative pipeline across Search, Gemini, Flow and Chrome.
Why this matters: if conversational video editing actually works at scale, world-model video generation becomes the first agentic media primitive. Watch for enterprise API access in Q3, and watch whether marketing teams, product designers and educators start adopting Omni inside Gemini Enterprise. That is the leading indicator of whether the demos translate into messy real workflows.
Read VentureBeat coverage →