Google I/O 2026: What Developers Should Build Now

Google I/O 2026 ran May 19-20 at Shoreline Amphitheatre. Two days and 100+ developer-facing announcements later, the real question is not what Google announced — it is what you should actually build with, what you should watch, what you can safely defer, and what you must migrate away from before a hard deadline. This priority matrix sorts every I/O release into one of four quadrants so engineering managers can finish their Q3 planning in one sitting.

The stakes are higher than a typical product keynote. Gemini 3.5 Flash went GA on May 19 with production pricing and a 1.05 million-token context window. Antigravity 2.0 shipped the same day with a new desktop app, a Go-based CLI, and an SDK that replaces the toolchain thousands of teams have been using since Antigravity 1.0 launched in November 2025. And buried in the developer keynote notes: Gemini CLI — the open-source tool with 100K+ GitHub stars — is being retired on June 18, 2026. That is 28 days from today.

This guide covers the full priority matrix, the Antigravity version distinction that most I/O coverage conflates, the AI Ultra pricing restructure, the open-source backlash over Gemini CLI, and the cross-vendor WebMCP story that almost every publication missed. Internal links connect to our sister-post deep dives for teams that want full technical detail on any quadrant item. For session-by-session coverage, see our Google I/O Day 2: 85-developer-session roundup.

The mental model is simple. ACT NOW means the API is GA or public preview with stable pricing and no waitlist — try it in production this week. WATCH means the surface is preview, beta, or "rolling out" — experiment in dev and plan a 3-month migration roadmap. WAIT means trusted testers only, research preview, or no announced developer surface — study the design pattern, do not build. HOLD means forced migration with a hard deadline, or no clear developer path at all — do not invest new resources, migrate out immediately.

Every item in Google's canonical 100-announcement index was assigned to a quadrant using three criteria: (1) is the developer API live and unrestricted today, (2) is pricing confirmed, and (3) is there a stable surface to build on without a waitlist. Where availability framing differed across sources, the Google I/O 2026 developer keynote round-up was treated as authoritative.

Eight developer-facing items from I/O 2026 cleared the ACT NOW bar as of May 19-20. They share three properties: the API or SDK is live without a waitlist, pricing is published, and Google's own documentation treats them as production-ready (even where the label says "preview," as with Managed Agents, the preview is fully open with stable pricing).

For detailed Gemini 3.5 Flash benchmarks, pricing tiers, and the full thinking_level API reference, see our dedicated Gemini 3.5 Flash benchmarks and API guide. The ACT NOW case is clear: this model is production-ready, cheaply priced, and already powering the surfaces that 8.5 million developers interact with monthly (per Pichai's May 19 stage number).

The WATCH quadrant is where most of the I/O excitement lives — and where most teams make their planning mistakes. These surfaces are real, they are coming, and some of them will be consequential. But they are not production-ready today, and a quarter spent building on a preview-stage API is a quarter that may need to be rebuilt.

The rule: set up a development-environment integration, track the changelog, and have a migration task card ready. Do not make a WATCH item a production dependency before it exits preview.

The WAIT quadrant contains the most hyped I/O announcements — including Gemini Spark, which generated more mainstream coverage than any other single announcement. The contrarian read is important here: excitement about a product is not the same as developer-readiness. Every item in WAIT either has no public developer API, is gated to a paid tier not broadly accessible, or has no announced SDK or developer surface.

The Spark framing deserves a specific note. Spark is architecturally interesting — it is an always-on personal agent running on dedicated Google Cloud VMs, built on the same Antigravity harness that powers the developer surfaces in ACT NOW. The platform play is Antigravity SDK. Spark is the consumer product that sits on top of it. Studying how Spark orchestrates context, manages long-horizon tasks, and integrates with Gmail and Calendar is valuable for any team designing always-on agent workflows. But building a product integration that depends on Spark's consumer-product surface — before a developer API exists — is building on a foundation that may never materialize.

The correct move: build on Antigravity SDK (ACT NOW) to implement the design patterns Spark demonstrates, and plan for a Spark API integration if and when Google opens the platform. Do not hold your Q3 roadmap waiting for a Spark SDK that has not been announced.

The HOLD quadrant has one item that demands immediate action from a large portion of the developer community: Gemini CLI is being retired on June 18, 2026. That is 28 days from this publication. Any team using gemini CLI in CI pipelines, developer onboarding documentation, GitHub Actions workflows, or local tooling needs to migrate within that window.

The migration is not zero-cost. Google's Transitioning Gemini CLI to Antigravity CLI post is explicit: "there won't be 1:1 feature parity right out of the gate." The migration path runs through antigravity migrate --from-gemini-cli, and the most critical features — Agent Skills, Hooks, Subagents, and Extensions (now called Antigravity plugins) — are documented in the migration guide but require testing.

Most I/O coverage conflates Antigravity 1.0 and 2.0. They are distinct products with different architectures, different model backends, and different release timelines. Getting this wrong leads to planning errors — particularly for teams currently running Antigravity 1.0 who are trying to understand what the "2.0" launch means for them.

The practical implication for teams currently on Antigravity 1.0: the May 23 migration tool handles the mechanical transfer of settings, keybindings, and extensions. There is no hard deadline to migrate from 1.0 to 2.0. The reason to migrate now is the model upgrade — Gemini 3.5 Flash running under Antigravity 2.0 is meaningfully more capable and cheaper than Gemini 3 Pro in Antigravity 1.0. For a detailed feature-surface comparison across the full Antigravity 2.0 stack, see our Antigravity 2.0 desktop, CLI, and SDK deep dive.

Google restructured its AI Ultra subscription at I/O, and the repricing has implications beyond the consumer product. For engineering managers evaluating team tooling costs, the new tiers define which Google developer surfaces are accessible.

The previous single AI Ultra tier was $250/mo. Google replaced it with two tiers: a new entry-Ultra at $100/mo (5× AI Pro limits, 20TB storage) and a repriced top-Ultra at $200/mo (down from $250, 20× AI Pro limits). Both Ultra tiers gate access to Spark. Most developer tools — Antigravity 2.0, the Gemini API, ADK 2.0, AI Studio — remain available outside the Ultra paywall on standard pay-per-use pricing. Ultra is primarily relevant as the Spark access tier and for teams that need 5-20× AI Pro limits on Gemini app usage.

For the full model pricing comparison across Gemini, Claude, and GPT as of May 2026, see our AI agent pricing landscape — May 2026 comparison. Our AI transformation engagements also include a procurement and tier-selection analysis as part of the platform assessment phase.

The Gemini CLI open-source community story is the most under-reported angle of I/O 2026. The Register and TechTimes both covered the developer community's reaction: Google accepted more than 6,000 merged pull requests to the open-source Gemini CLI repository — building the tool to 100K+ GitHub stars and a substantial community of contributors — and then announced that its replacement, Antigravity CLI, is not open-source for non-enterprise users.

The developer-community interpretation is straightforward: Google harvested OSS community labor to build a capable tool, then closed the gate at the enterprise boundary. Whether that reading is fair or not, it is the reaction that is circulating on Hacker News and in engineering Slack channels, and it has practical implications for teams making toolchain decisions.

The analytical framing that matters for planning: Antigravity CLI is a superior technical product to Gemini CLI in terms of capability and integration with the 2.0 stack. But it is a closed-source binary from a single vendor, not a community-maintained project with the ability to fork, audit, or self-host. Teams with security, compliance, or vendor-dependency requirements should factor this into their June 18 migration decision. The alternative — maintaining a Gemini Code Assist Standard or Enterprise license to preserve CLI access — keeps the familiar toolchain but at a meaningfully higher cost.

WebMCP was announced quietly in the Chrome at I/O 2026 — 15 updates post. Most mainstream I/O coverage did not flag it. The significance is structural: WebMCP is a new open web standard that lets browser-based AI agents call structured tools — JavaScript functions, HTML form fields, registered web APIs — instead of simulating mouse clicks and DOM scraping. The origin trial opens in Chrome 149.

What most coverage missed: WebMCP is being co-developed by Google and Microsoft inside the W3C WML Community Group. That cross-vendor alignment from day one is rare for a browser feature — it means WebMCP is tracking toward cross-browser support rather than being a Chrome-only proprietary surface. The last browser API to launch with simultaneous Google-Microsoft co-development was the Progressive Web App spec — and that became the foundation for how modern web apps are installed and cached.

The practical implication for teams building browser-embedded AI agents today: WebMCP is months away from cross-browser GA, but the architecture you choose now will determine how much rework the migration costs. Building browser agent interactions as structured function calls — rather than click simulations — already aligns with what WebMCP will require. Start treating DOM-simulation as technical debt now, not when WebMCP lands.

For teams building SEO and search-integration tools, the WebMCP story connects to the broader Generative UI and AI Mode surface changes from I/O. Our complete Google I/O 2026 AI announcement guide covers the full search and SEO surface changes, including Generative UI in Search and the Universal Cart rollout.