7 Things From Google I/O 2026 That Matter to .NET Developers (And 4 You Can Skip)

I watched the whole keynote at 2x. Here’s what actually changes for your C# stack — and what’s just theater for someone else’s pipeline.

I watched the entire I/O 2026 keynote at 2x speed, took notes, and sat down to write up what mattered to .NET developers. The first draft had three bullet points.

Read for free

So I went back, watched the developer keynote too, and combed through Google’s own “100 things we announced” post. The list grew. Not all of it matters to you — most of it ships for someone working in a different stack than yours, building for a different platform than yours, paying for a different cloud than yours.

But some of it absolutely does. Here’s the .NET-shaped filter.

What matters

1. The Google Gen AI .NET SDK quietly became first-class

This one didn’t ship at I/O — it shipped in October 2025, a few weeks before .NET 10’s stable release, and most .NET developers missed it. Google now maintains a unified Gen AI SDK with first-party support for Python, Go, Node.js, Java, and C#. Mete Atamel announced it on the Google Cloud blog.

Until then, calling Gemini from a C# app meant either the excellent-but-unofficial Mscc.GenerativeAI NuGet package or hand-rolled REST calls with HttpClient. Both worked. Neither was the experience the Python folks were getting.

// The new official path
using Google.GenAI;

var client = new GoogleGenAI(apiKey: Environment.GetEnvironmentVariable("GEMINI_API_KEY"));
var response = await client.Models.GenerateContentAsync(
    model: "gemini-3.5-flash",
    contents: "Explain why EF Core's IQueryable beats IEnumerable for SQL filters."
);
Console.WriteLine(response.Text);

The implication: if you’ve been picking between Azure OpenAI and Anthropic over REST for your AI features, Gemini is now a real third option without the SDK-shaped friction. Worth one evening of eval work before your next sprint.

2. Gemini 3.5 Flash is the new Pro-killer

Google’s framing was loud, but the numbers back it. Gemini 3.5 Flash outperforms Gemini 3.1 Pro on coding and agentic benchmarks while running roughly four times faster than competing frontier models. It’s generally available today via the Gemini API, Antigravity, and Android Studio.

Why a .NET developer cares: if your production AI features run on the Pro tier of any vendor — chat, function calling, RAG retrieval reranking, code analysis — you’re probably overpaying for intelligence you’re not using. Flash tier models from every major vendor are now close enough to Pro tier on most real workloads that the cost gap is harder to justify.

Run your eval set against Gemini 3.5 Flash this week. If your tests pass within a 5% margin of your current Pro-tier model, you have a budget conversation to have.

3. Antigravity 2.0 — and the .NET solution problem that hasn’t gone away

Google didn’t just refresh Antigravity. They rewrote it into a five-surface platform: a VS Code-fork desktop app with multi-agent orchestration, a CLI rewritten in Go (invoked as agy), an SDK, a Managed Agents API tier, and an enterprise deployment path. All powered by Gemini 3.5 Flash. The platform scores 76.2% on SWE-bench Verified — about one percentage point behind Claude Sonnet 4.5's top score.

Read that benchmark with a grain of salt. SWE-bench measures Python tasks against open-source repos. .NET solutions don’t show up.

Here’s where the demo magic frays. Agent-first coding tools — Cursor, Claude Code, Copilot in agent mode, and now Antigravity — still don’t grok multi-project .NET solutions natively. They handle a single .csproj beautifully. Hand them a 14-project .sln with a domain assembly, three feature assemblies, a contracts library, an EF Core data project, and a couple of test projects, and you'll watch them confidently edit the wrong project, miss source-generator hooks, or implement an interface in the consumer instead of the contract assembly.

Whether Antigravity 2.0 has materially fixed this for .NET, nobody outside Google knows yet. Try it on a real solution, not a single-project sample. Watch what it does when it needs to add a method to an interface that lives three projects away from the implementation. That’s the test that separates “agent that codes” from “agent that codes in your codebase.”

4. Managed Agents landed in the Gemini API

This is the announcement that most .NET shops should read twice. Google now offers a Managed Agents tier inside the Gemini API — a single API call spins up an isolated, persistent agent environment with code execution, file management, web browsing, and reasoning, all running in Google’s sandboxed Linux.

In Microsoft’s world, the equivalent stack is Semantic Kernel agents plus Microsoft Agent Framework plus your own orchestration code plus your own sandbox plus your own state management. Five components you maintain. Google compressed that into one API endpoint.

The trade is the standard one: less control, faster setup, deeper lock-in. If you’re three sprints into building agent workflows in Semantic Kernel, this doesn’t kill your work. But if you’re scoping a new agent feature next quarter, a one-day spike comparing Managed Agents to your current Semantic Kernel design is worth the time. The cost calculation might surprise you.

5. Gemini CLI is dead. You have weeks to migrate.

Google retired Gemini CLI on the day Antigravity 2.0 shipped. The official shutdown date for consumer users is June 18, 2026 — about three weeks from now.

If you’ve automated anything against gemini-cli — PR review bots, release-note generation, commit message linting, CI/CD steps — open that file and start porting. The replacement is the Antigravity CLI (agy), built in Go.

Two things the migration guide won’t lead with:

• The new CLI is a closed binary. Gemini CLI was Apache 2.0 licensed on GitHub. Antigravity CLI is not. If your shop has open-source-only tooling policies, this becomes a procurement problem before it’s a technical one.
• The /usage command only refreshes after you quit and relaunch. You cannot monitor quota during a long agent run. Multiple developers on the Google AI Developers Forum have already reported running out of quota mid-task. Plan accordingly.

6. SynthID + C2PA Content Credentials are becoming table stakes

Google expanded its watermarking program to include C2PA Content Credentials. Translation: every image generated by Imagen, every video generated by Veo, every audio clip from Gemini’s audio models now carries a verifiable provenance signature that survives most edits.

You’re probably wondering why a .NET developer should care. Two reasons.

First, if your .NET app generates any AI media — even if you’re calling Imagen via the new SDK, or running Stable Diffusion through ML.NET, or piping Veo output into a content pipeline — your B2B customers are about to start asking about provenance. The audit-trail expectation is shifting from “nice to have” to “required for enterprise procurement” inside the next 12 months.

Second, Microsoft will match this within a quarter or two. When Azure OpenAI ships its own C2PA layer, you’ll want to have already thought through where in your pipeline the watermarks get checked, who’s allowed to strip them, and how your storage layer preserves them.

7. The Antigravity SDK is callable from outside the Antigravity app

Buried in the developer keynote: the Antigravity SDK exposes the same agent primitives that power the desktop app and the CLI — code execution, file management, web browsing, reasoning — all running in Google’s secure Linux sandbox.

The interesting part: you don’t need to use Antigravity to use the SDK. You can build a custom agent in .NET — Semantic Kernel, your own framework, whatever — and call out to the Antigravity SDK for the hard parts. Sandboxed code execution. Browser-based UI verification. File manipulation in an isolated environment.

That’s a different design pattern from “rewrite everything in Antigravity.” It’s the same pattern .NET teams already use with Azure Functions for sandboxed compute — you just have a new option for the runtime. Worth a prototype before you commit to one platform.

What you can skip

Four announcements that ate keynote time and won’t move your stack:

Gemini Spark and Daily Brief. Personal AI agent, 24/7 proactive assistance. A consumer feature with a Google account login. Cool product. Nothing for you to integrate against.

Universal Cart and agentic commerce. Genuinely interesting if you sell merchandise through Google’s ecosystem. Otherwise: skip until 2027 at the earliest. The Agent Payments Protocol might matter eventually, but the .NET tooling story is years away.

Samsung Intelligent Eyewear. Fall 2026 hardware launch. Maps, texts, photos, no .NET SDK story, no developer story beyond “wait and see.”

Google Pics, voice in Gmail, voice in Docs, voice in Keep. End-user Workspace features. Useful if you live in Google Workspace. Irrelevant if you’re shipping .NET applications.

The bigger pattern

Strip away the consumer demos and I/O 2026 was Google fighting OpenAI and Anthropic head-on for the developer mindshare that used to be theirs by default. The Antigravity rewrite is the proof — they understand that the agent-first IDE is now the battleground, and Claude Code and Cursor have a head start they need to close.

The wins for .NET developers are quiet ones. An official SDK that should have shipped a year ago. A Flash-tier model worth re-running your evals against. A more aggressive answer to Cursor and Claude Code that may or may not work on real .NET solutions yet. A heads-up that Gemini CLI is dying in three weeks.

Three of those seven things change something concrete in your toolchain this month. That’s worth the 90 minutes of keynote you didn’t have to watch.