Google I/O 2026: What Gemini 3.5 Actually Means

Google I/O 2026 opens in 48 hours — and for the first time in the conference's history, the headline launch is expected to be a Flash-tier model rather than a Pro. Gemini 3.5 Flash is rumored to hit the developer keynote on Tuesday May 19 with a stable gemini-3.5-flash model ID, leaked pricing of $1.50/$9.00 per million tokens, and a new thinking_level API surface replacing the old integer budget.

The inversion matters. Google's usual cadence has always been Pro first, Flash as the trimmed-down follow-up. Shipping 3.5 Flash before 3.5 Pro is a deliberate “small models win developer mindshare” narrative — and it lands in the same week that Anthropic opens Code with Claude in London. The May 19-20 window is the single most compressed AI-news period of the quarter.

This guide separates confirmed agenda items from pre-keynote leaks from pure analyst speculation, deep-dives the rumored Gemini 3.5 Flash API surface and Antigravity 2.0 expectations, and tells you exactly what to pre-bookmark before the stream goes live Monday morning.

Google I/O 2026 runs Tuesday-Wednesday May 19-20, with a Monday May 18 community pre-event. The opening keynote is delivered by Sundar Pichai at 10:00 AM PT from the Shoreline Amphitheatre in Mountain View — livestreamed free on YouTube. The developer keynote follows immediately, led by Josh Woodward (VP, Google Labs and Gemini app) and Hema Budaraju (VP, AI Mode in Search). Both are on the confirmed agenda per io.google/2026.

The breakout session library opens to logged-in Google accounts immediately after the opening keynote. Android 17 developer sessions — covering adaptive layouts, predictive back, and the new Foreground Services API — are confirmed on the breakout track. So are Gemini-in-Chrome sessions and the AI Mode in Search GA expansion track.

Most pre-I/O coverage mashes confirmed agenda items together with leaks and analyst conjecture. The table below applies a three-tier taxonomy — each row is tagged by evidence quality. CONFIRMED means it appears on the official io.google/2026 schedule. LEAKED means a named publication with direct sourcing has reported it as pre-keynote intelligence. SPECULATION means analyst or market-watcher inference with no named source.

The takeaway from the triage: Gemini 3.5 Flash and Antigravity 2.0 have multi-outlet named sourcing and are the closest to certainty among unconfirmed items. Universal Cart and Gemini Omni are plausible given Google's public product roadmap but lack primary source corroboration as of today. Track the official blog.google “Everything announced at I/O” post — it typically goes live within an hour of the opening keynote closing.

Pre-keynote leaks from AI Studio changelog feeds point to a new API surface for Gemini 3.5 Flash — the most developer-relevant change is the replacement of the thinking_budget integer with a string enum called thinking_level. The reported values are minimal, low, medium, and high, with mediumas the default. This mirrors the design philosophy already present in the Claude API's thinking block, which Anthropic shipped earlier this year.

The existing Gemini model IDs context: the current Flash tier sits at Gemini 3 Flash (launched Dec 17, 2025) and Gemini 3.1 Flash-Lite (GA May 7-8, 2026). Gemini 3.5 Flash is rumored to leapfrog both without replacing 3.1 Flash-Lite in the cost tier — they would coexist. The Gemini 3.1 Flash-Lite cost-tier breakdown documents that existing hierarchy before 3.5 changes the picture.

On pricing, the leaked $1.50 input / $9.00 output per million tokens positions Gemini 3.5 Flash between the current 3 Flash ($0.30 in / $2.50 out) and Gemini 3.1 Pro ($4.50 in / $18.00 out). That gap is where agentic workloads — multiple sequential tool calls, long context chains, structured reasoning loops — currently sit without a well-priced option. If the leak holds, 3.5 Flash fills it directly. The verified Gemini 3.5 Flash benchmark guide will publish once confirmed specs are public after Tuesday's keynote.

Antigravity 1.0 launched November 18, 2025 as a VS Code fork with Manager View running five parallel agents, Chrome integration, and Gemini 3 Pro as the native model — hitting 76.2% on SWE-Bench at launch as a free preview. Our Antigravity 1.0 deep-dive in the AI-IDE landscape documented that baseline in detail.

Antigravity 2.0 is expected to ship on Gemini 3.5 Flash rather than a Pro-tier model — which is itself part of the Flash-first narrative. The leak profile from The Verge and TechCrunch points to four additions over the 1.0 baseline: a native desktop app (not a VS Code fork), a CLI surface, a first-party SDK, and scheduled tasks with slash-command support.

Hema Budaraju's session is the one that SEOs and search strategists should watch most carefully. Pre-event analyst notes from Search Engine Land expect the session to frame AI Mode as now serving over a billion people monthly — a milestone that would make it one of the fastest growth curves in Google Search's history. That figure remains a pre-keynote expectation, not a self-published Google number as of this writing.

What the confirmed session content tells us: AI Mode in Search GA expansion is on the breakout track. That means moving beyond the limited rollout into broader availability, likely U.S.-wide or global, depending on regulatory status. The practical implication for digital marketing teams is that the organic click distribution patterns measured in 2025 may shift again — AI Mode surfaces structured summaries rather than ten blue links, rewarding content that answers intent directly rather than optimizing for position. Our agentic SEO practice is structured around exactly this shift.

Beyond GA expansion, analyst notes speculate about intent anticipation — a feature that surfaces results or actions before the user completes a query, modeled on behavioral patterns across Search, Gmail, and Assistant. That remains in the speculation tier as of May 17; watch Budaraju's session for any mention of “proactive Search” framing.

Two surfaces are in the leaked-but-expected tier and are worth watching separately. Gemini Spark is the consumer story; the Managed Agents API is the developer infrastructure story.

The Managed Agents API, if it ships as speculated, would place Google in direct competition with Anthropic's agent runtime and OpenAI's Responses API at the infrastructure layer. The three labs are converging on a similar primitive: a managed execution environment where reasoning, tool calls, and code execution are bundled together rather than left to developer scaffolding. For teams already building on the agentic coding agent landscape, this represents a shift from DIY orchestration to platform-managed agent infrastructure.

Google's historical cadence has been consistent: ship the Pro tier first, establishing the quality ceiling, then release Flash as the efficient derivative 8-16 weeks later. Every Gemini generation from 1.0 through 3.1 followed this pattern. Gemini 3.5 breaks it — Flash ships first, Pro is reportedly held for June 2026.

This is not an accident. It is a deliberate signal that the Flash-Pro intelligence gap has narrowed to the point where Flash can lead the narrative. The Flash-first move makes sense only if Google believes 3.5 Flash is meaningfully better on agentic and coding benchmarks than 3.1 Pro — because that is the audience the developer keynote is playing to. Ship Flash first, own developer mindshare with the affordable, fast model, then ship Pro to the enterprise and research audience next month.

The forward projection: if 3.5 establishes Flash-first as the new pattern, expect every subsequent Google model generation to lead with the Flash tier. That reframes how developer teams should track Google's model roadmap — the Flash launch date becomes the primary planning horizon, not the Pro launch date. Teams currently building on where Gemini 3.1 Pro sits today should pre-budget for a 3.5 Flash evaluation cycle starting Tuesday.

The horizontal bars above show every generation in the same order — Pro then Flash — until 3.5. The inversion is not subtle. Pre-event analysis from Stratechery has framed this as Google shifting developer-acquisition strategy: instead of winning with the most powerful model, win with the most affordable model that is still capable enough for production agentic workloads. Whether that framing lands depends on what the actual 3.5 Flash benchmark numbers show — and those won't be available until the keynote on Tuesday.

The May 19-20 window is unusual even by the standards of the compressed 2026 AI calendar. Three major events run in parallel — and the developer audience has to choose where to point its attention.

The counter-programming collision is not coincidental at the industry level — major conferences cluster around the same windows because enterprise calendar constraints make any other scheduling harder. But for developers tracking all three spaces, the practical move is to watch the Google I/O opening keynote live on Tuesday morning, catch the Code with Claude London sessions in the afternoon archive (Anthropic typically releases session recordings within hours), and scan the SAP Sapphire enterprise AI track for any partner announcements that affect your stack.

The Code with Claude London preview covers what Anthropic is expected to ship in that window. The two previews are designed to be read together as complementary context for the same 48-hour period.

The developer keynote moves fast — announced model IDs, pricing tables, and API surfaces go from slide to live documentation in minutes. Developers who are already oriented land in a much better position to evaluate quickly. Here is what to set up before Tuesday.