Google just dropped its biggest AI release of the year. At I/O 2026, the company announced Gemini 3.5 Flash (available today), a multimodal video platform called Omni, background agents called Spark, and a full agent operating system called Antigravity 2.0. If you run a service business and you’ve been building on OpenAI or Anthropic, this changes the math.
This isn’t a research preview or a staged rollout. Gemini 3.5 Flash is GA right now across the Gemini app, API, AI Studio, and enterprise surfaces. Google is clearly done playing catch-up. They’re making a production play.
What happened
- Gemini 3.5 Flash went live immediately. It’s Google’s fastest and strongest agentic/coding model, with a 1M token context window, 65k max output tokens, and four thinking levels (minimal, low, medium, high). It also supports “thought preservation” across multi-turn conversations.
- Gemini Omni launched for video. It merges Gemini’s reasoning with generative media capabilities. You can feed it text, images, audio, or video and get video edits or generation back. Available now for paid Gemini users, rolling out to YouTube Shorts this week, with APIs coming soon.
- Antigravity 2.0 became a full agent platform. Desktop app, CLI, SDK, and Managed Agents in the Gemini API. One API call gives you an agent plus a hosted Linux sandbox with Bash, Python, Node, file access, and browsing.
- Spark brings background agents. These run on dedicated Google Cloud VMs, meaning they keep working even after you close your laptop.
- Google showed off a demo where Antigravity and Gemini 3.5 Flash built a functioning OS in 12 hours using 93 parallel sub-agents, 15,000+ model requests, and 2.6 billion tokens for under $1,000 in API credits.
The numbers
- 900M+ monthly users on the Gemini app, available in 230+ countries and 70+ languages.
- 3.2 quadrillion tokens/month processed by Google, up 7x year over year from 480 trillion.
- Pricing: $1.50 per 1M input tokens, $9.00 per 1M output tokens, with a 90% discount on cached input.
- Speed: Google claims 4x faster than comparable frontier models, up to 12x faster inside Antigravity. Independent testing from Artificial Analysis clocked it at over 280 output tokens per second.
- Benchmarks: Terminal-Bench 2.1 at 76.2%, GDPval-AA Elo of 1656, MCP Atlas at 83.6%, MMMU-Pro at 84%. Artificial Analysis gave it an Intelligence Index of 55 (up 9 points over Gemini 3 Flash).
- Arena ranking: #9 overall in Text Arena and #9 in Code Arena: Frontend, with a score of 1507 (a 70-point jump over Gemini 3 Flash).
- The cost trade-off: Artificial Analysis found 3.5 Flash is 5.5x costlier to run than Gemini 3 Flash and 75% costlier than Gemini 3.1 Pro. Flash is no longer the budget option it used to be.
5 things service business operators should pay attention to
- “Flash” doesn’t mean cheap anymore. Google is using the Flash label for what used to be Pro-level capability. The performance is real, but so is the price bump. Budget your AI costs accordingly.
- Background agents are a big deal. Spark agents running on cloud VMs can handle long-running tasks without a human babysitting them. Think overnight document processing, multi-step research, or batch workflows that used to require custom infrastructure.
- Video generation just got an API. Omni’s developer APIs are coming in weeks. If you serve clients who need video content (marketing agencies, real estate firms, ecommerce brands), this is worth watching closely.
- The agent tooling is production-ready. Managed Agents in the Gemini API give you a sandbox with code execution, browsing, and file access in a single API call. That’s a massive shortcut for building agent-based services.
- Multi-agent orchestration is the new default. Google’s marquee demo used 93 parallel sub-agents. They’re clearly betting that the future isn’t one big model call. It’s many small coordinated ones. If your workflows still rely on single-prompt chains, you’re leaving performance on the table.
The hot take
Google just made the strongest case that the AI platform war isn’t about having the smartest model. It’s about having the fastest model, the deepest integrations, and the execution layer that lets agents actually do things. Gemini 3.5 Flash probably isn’t the smartest model on the market right now. Some benchmarks are mediocre, and skeptics have raised fair points about pricing relative to GPT-5.5-medium. But none of that matters if Google is the only company shipping a complete stack where your agent can spin up a VM, run code, browse the web, call sub-agents, and keep working while you sleep. OpenAI and Anthropic should be worried. Not about benchmarks. About plumbing.
The Agency OS play
If you run any kind of service business, this week’s move is simple: get API access to Gemini 3.5 Flash and test it against whatever you’re currently running on Claude or GPT. Specifically, test it on your longest, most token-heavy workflows. The 1M context window and 65k output limit make it a strong candidate for document-heavy tasks like contract review, financial report analysis, or intake processing. The speed advantage (4x faster than comparable models, per Google) means your agents finish faster and your clients wait less.
Second, look at Managed Agents in the Gemini API. If you’ve been stitching together your own agent infrastructure with custom sandboxes and tool-calling logic, Google just gave you a hosted version for free (well, for the cost of tokens). Try building one workflow end to end using a Managed Agent. Pick something repetitive that your team does manually: pulling data from multiple sources, generating a report, formatting deliverables. See if a single API call can replace your current multi-step pipeline.
Third, if you serve clients who care about video or visual content, bookmark Omni’s API launch. When those APIs go live in the coming weeks, the ability to generate and edit video from text or image inputs will open up service offerings that were previously out of reach for small shops. Don’t wait for the launch to start thinking about what you’d build. Sketch the workflow now so you’re ready to prototype on day one.