Google disclosed $190 billion in annualized AI capital expenditure at Google I/O on May 20, 2026. That figure, larger than the GDP of 150 countries, explains more about what Gemini has become than any benchmark score could. The company that spent 2023 scrambling to match OpenAI has built something its competitors cannot easily replicate: a platform spanning consumer devices, developer tools, enterprise infrastructure, and an autonomous agent ecosystem, all powered by custom TPU silicon at a scale no rival matches.
This guide covers where Gemini stands right now, what changed at I/O 2026, and how Google’s AI platform compares to ChatGPT and Claude as of May 2026.
The Model Stack
Google’s model lineup has changed dramatically from the Gemini 1.5 and 2.0 era. The naming convention now follows version numbers with a speed/quality split rather than the old Ultra/Pro/Flash capability tiers.
Gemini 3.5 Flash launched at I/O 2026 on May 19 and is now the default model in the Gemini app. It outperforms the previous flagship (Gemini 3.1 Pro) on coding and agentic benchmarks while running four times faster than competing frontier models. On Terminal-Bench 2.1, it scores 76.2%. On MCP Atlas, the standard agentic tool-use benchmark, it reaches 83.6%. API pricing: $1.50 per million input tokens, $9.00 per million output tokens.
Gemini 3.1 Pro remains the top-tier model for complex, multi-step reasoning. It powers the most demanding tasks in the Gemini app and Antigravity development environment. API pricing runs $2/$12 per million input/output tokens up to 200K context, doubling above that threshold.
Gemini 3 Flash and Flash-Lite serve high-volume, cost-sensitive production workloads at $0.50/$3.00 and $0.25/$1.50 per million tokens respectively.
Gemini 3.5 Pro is running internally at Google and scheduled for public release next month. Based on 3.5 Flash’s performance trajectory, expect it to reset benchmarks for the full-size model tier.
Gemini Omni, also announced at I/O, combines Gemini’s reasoning capabilities with Google’s generative media models (Veo, Imagen) into a unified system that produces any output from any input. It starts with video generation and editing, creating a single model that can reason about content and then produce it.
The context window across the 3.x series holds at one million tokens. In practical terms, that is approximately 1,500 hours of audio, 3,000 pages of dense text, or a complete mid-sized software codebase loaded in a single pass. Claude from Anthropic now matches this with a 1M context window on Opus 4.7, while GPT-5.5 tops out at 256K tokens.
Beyond raw context length, Gemini’s multimodal processing remains a genuine differentiator. The 3.x models natively handle text, images, audio files, video (including hours of footage with frame-level understanding), PDFs, and code. The video analysis capability is particularly distinct: upload a recorded meeting, a product demo, or a sports broadcast, and Gemini can answer specific questions about moments, identify patterns across frames, and produce structured breakdowns. No other frontier model handles long-form video with comparable depth.
Gemini Spark: The Always-On Agent
The most significant announcement at I/O 2026 was not a model release. It was Gemini Spark, a 24/7 personal AI agent that takes actions on your behalf across your entire digital life.
Spark runs on dedicated cloud virtual machines, meaning it operates continuously even when your phone or laptop is powered off. It handles multi-step tasks in the background: scanning apartment listings to your specifications, monitoring sneaker drops from specific athletes, booking appointments through Chrome’s auto-browse capability, and executing cross-app workflows like transferring a grocery list from a notes app into a delivery order.
Google is rolling Spark out first to AI Ultra subscribers in the U.S., with broader availability planned for later in 2026. The system requires explicit user approval for high-stakes actions (purchases, account changes, outbound messages), and progress displays through Android Halo, a new lock-screen interface showing live agent activity and status updates.
This puts Google in direct competition with OpenAI’s Operator and Anthropic’s Computer Use. The architectural differences matter: Operator runs inside a browser sandbox, Computer Use controls a desktop session that must stay active, and Spark runs persistently on cloud infrastructure independent of any local device. Spark’s model is the only one that works while the user sleeps. For a broader look at how agents fit into the software landscape, see our guide to what AI agents actually are.
Gemini Intelligence on Android
Separate from Spark, Gemini Intelligence integrates AI capabilities directly into the Android operating system. Announced in mid-May 2026, it transforms every compatible device into an AI-native platform where intelligence works across the entire OS rather than being confined to a single app.
Features rolling out this summer on select Samsung Galaxy and Google Pixel devices:
- Cross-app task execution: Gemini completes workflows spanning multiple apps (ride booking, grocery orders, research synthesis) without requiring users to switch between interfaces
- Chrome AI browsing: In-browser research, page summarization, price comparison, and automated repetitive web tasks including form filling and parking reservations
- Rambler: Converts natural, unstructured spoken thoughts into polished, contextually appropriate written messages
- Natural-language widgets: Users create custom home-screen widgets by describing what information they want in plain English
With over 3 billion active Android devices worldwide, Gemini Intelligence reaches users who will never install a standalone AI app or configure an API key. OpenAI’s planned AI phone is still at least a year from production. Apple Intelligence exists but has not demonstrated comparable agentic capabilities on device. Google has the distribution layer locked down in a way no other AI lab can currently match.
Antigravity 2.0: The Developer Stack
Google launched Antigravity 2.0 at I/O as its unified, agent-first development platform. The package includes a standalone desktop application for agent orchestration, a CLI, an SDK, and Managed Agents in the Gemini API.
The Managed Agents API is particularly notable: with a single API call, developers can spin up an agent that reasons, uses tools, and executes code inside an isolated Linux environment. Dynamic subagents enable parallelized workflows. Scheduled tasks handle background automation. All of it defaults to Gemini 3.5 Flash as the reasoning engine.
For enterprise deployments, the Gemini Enterprise Agent Platform adds security controls, compliance tooling, and governance infrastructure on top of the same foundation. Google AI Studio remains the free prototyping entry point.
This positions Antigravity alongside Anthropic’s Model Context Protocol and OpenAI’s Assistants API as the three major agent development stacks competing for developer adoption in 2026. Google’s structural advantage: Antigravity agents can access Workspace data natively, run on Google Cloud infrastructure, and deploy to Android devices through a single integrated pipeline.
Consumer Pricing: Four Tiers
Google restructured its consumer AI subscriptions at I/O 2026:
| Plan | Monthly Price | Key Features |
|---|---|---|
| Free | $0 | Gemini app with daily limits, Gemini 3 Flash |
| AI Plus | $7.99 | Higher usage limits, Gemini 3.5 Flash access |
| AI Pro | $19.99 | Full model access (3.1 Pro), Antigravity, 2TB storage |
| AI Ultra | $99.99 | 5x Pro limits, priority Antigravity, Spark agent, 20TB, YouTube Premium |
A second Ultra tier at $200/month offers 20x Pro usage limits for power users and teams. The billing model itself is changing: Google is shifting from simple daily prompt caps to a compute-used system that factors prompt complexity, with allowances refreshing every five hours until a weekly ceiling.
For comparison, ChatGPT Plus runs $20/month, Claude Pro costs $20/month, and enterprise tiers from both OpenAI and Anthropic start around $60 per seat. Google’s tiered approach gives it price coverage from $7.99 for casual users to $200 for power users, a range neither competitor currently matches at the consumer level.
Gemini vs. ChatGPT vs. Claude: May 2026
| Capability | Gemini | ChatGPT (GPT-5.5) | Claude (Opus 4.7) |
|---|---|---|---|
| Top reasoning model | 3.1 Pro (3.5 Pro next month) | GPT-5.5 | Opus 4.7 |
| Speed model | 3.5 Flash (4x faster) | GPT-5.5 Mini | Sonnet 4.6 |
| Context window | 1M tokens | 256K tokens | 1M tokens |
| Agent platform | Spark + Intelligence | Operator + Deep Research | Computer Use + Cowork + Code |
| Developer tools | Antigravity 2.0 | Assistants API | MCP ecosystem |
| Device reach | Android (3B+ devices) | Planned phone (2027) | Desktop apps |
| Entry price | $7.99/mo | $20/mo | $20/mo |
Choose Gemini if you already live in Google’s ecosystem (Workspace, Android, Chrome), need the deepest multimodal capabilities (especially long-form video), want an affordable entry tier, or need an always-on cloud agent through Spark.
Choose ChatGPT if you want the broadest third-party plugin ecosystem, rely on GPT-5.5’s agentic workflow capabilities for complex multi-step tasks, or need tight integration with Microsoft 365 through Copilot.
Choose Claude if you prioritize careful, extended reasoning over raw speed, need best-in-class performance on coding and analytical tasks, or want to build agents on Anthropic’s open MCP protocol.
The Full-Stack Bet
Google’s strategy is structurally distinct from every other AI company because no one else owns the entire vertical: custom silicon (TPU v6), frontier models (Gemini 3.5), consumer applications (Workspace, Search, Android), a developer platform (Antigravity), and device distribution reaching 3 billion users. OpenAI and Anthropic build exceptional models but depend on partners for infrastructure and have no device layer. Apple has the devices but not the models. Microsoft has enterprise distribution but licenses its AI from OpenAI.
The Gemini platform of May 2026 is unrecognizable from the Bard prototype that launched to mixed reviews in early 2023. Whether it becomes the default AI for most of the world depends less on any single benchmark and more on whether Spark, Intelligence, and Antigravity deliver on the promise of AI that acts autonomously rather than just responding to prompts. Based on what Google showed at I/O 2026 and the $190 billion in infrastructure backing it, the early results suggest Google is closer to that vision than most analysts expected twelve months ago.