Google I/O 2026: Gemini 3.5 Flash, Antigravity 2.0 and Managed Agents Go Live

Google's annual developer conference on 19–20 May 2026 was framed almost entirely around agents. The centrepiece was Gemini 3.5 Flash, now generally available through the Gemini API, AI Studio, Android Studio and Google's Antigravity platform. Google said the model beats its own Gemini 3.1 Pro on coding benchmarks while remaining in the fast, lower-cost tier; third-party trackers list pricing at $1.50 per million input tokens and $9 per million output, with a context window above one million tokens. A larger Gemini 3.5 Pro was described as in testing and due the following month, and a new "Omni" family for any-to-any multimodal generation debuted with Omni Flash for subscribers.

On the platform side, Antigravity 2.0 arrived as a desktop application, CLI and SDK for building and hosting custom agents, and a new Managed Agents capability in the Gemini API lets a developer provision a remote Linux environment, with sandboxing and web browsing, in a single call. Google also previewed WebMCP, a proposed open standard for websites to expose tools directly to browser-based agents, and shipped Gemini Spark, a background personal agent for phones and laptops, to Ultra subscribers in beta. Consumer-facing numbers were large: AI Mode in Search passed one billion monthly users and now defaults to Gemini 3.5 Flash, while a new Google AI Ultra tier is priced at $100 per month.

Two provenance announcements deserve attention from governance teams. SynthID watermark verification is now built into Search and Chrome, and Google said the detector has been used 50 million times. C2PA content credentials are being surfaced in the Gemini app, Search and Chrome, which brings a cross-industry standard to some of the most widely used consumer surfaces.

The context is competitive and regulatory at once. Anthropic used the same day to launch self-hosted sandboxes that keep agent execution inside a customer's own cloud, and Microsoft had made computer-use agents generally available a week earlier. Every major vendor is now selling hosted agent runtimes, which shifts the security perimeter from the model to the environment the model can act in. Meanwhile the EU's transparency deadline for AI-generated content was just pulled forward to 2 December 2026, making watermarking and provenance tooling a compliance need rather than a nice-to-have.

What it means for leaders

• Decide where agents run before deciding which model. Managed Agents, Antigravity hosting and rivals' self-hosted sandboxes represent different data-residency and audit postures; procurement should compare isolation, logging and credential handling, not just benchmarks.
• Extend identity and access controls to agents. A remote Linux environment that browses the web on your behalf needs its own service identity, scoped secrets and egress policy, consistent with NIST AI RMF "Manage" guidance on third-party components.
• Plan for provenance obligations now. SynthID and C2PA support in mainstream browsers makes it feasible to verify content; organisations publishing synthetic media in the EU should adopt one or both ahead of the December 2026 deadline.
• Watch WebMCP. If websites begin exposing tools to agents, corporate web properties will need policies on what actions an automated visitor may take, and security teams will need to test them.
• Reassess model-tier assumptions. A Flash-class model outperforming last generation's Pro changes cost models for agentic workloads; revisit usage forecasts and any contractual minimums tied to premium tiers.