Google I/O 2026 kicks off May 19-20 at Shoreline Amphitheater, and this year's conference arrives at a pivotal moment. Google Cloud posted roughly 50% year-over-year growth in Q1 2026 — driven almost entirely by Gemini API consumption and TPU inference economics. Google isn't catching up on AI anymore. It's pressing an advantage.
Here's what to expect and why it matters for anyone building with AI models.
Gemini 3.1 Ultra: The Headline Announcement
Gemini 3.1 Flash-Lite is already in production at $0.25 per million tokens with a 1 million token context window. The I/O keynote will almost certainly unveil Gemini 3.1 Ultra — the frontier model in the Gemini 3 family.
What the signals say
- Extended context window: Gemini 3.1 Pro already supports 1 million tokens. Ultra is expected to push toward 2-4 million tokens. A 4M context window can hold approximately 3,000 pages of text or an entire mid-size codebase. This changes what's possible for long-context reasoning, legal document analysis, and full-repository code review.
- Native multimodal improvement: Project Astra (Google's multimodal agent) has been in preview. Gemini 3.1 Ultra is expected to bring Astra's real-time audio and video understanding into production API access — live video processing, real-time transcription with contextual understanding, and multi-turn visual conversations without a separate vision model call.
- Coding benchmarks: Google has been aggressive on SWE-bench. Gemini 3.1 Ultra is expected to push scores above 75%, putting it in direct competition with Claude Opus 4.7 (64.3% on SWE-bench Pro) and whatever OpenAI ships next.
Pricing speculation
If Google follows its pattern, Gemini 3.1 Ultra will be priced competitively against GPT-5.5 and Claude Opus 4.7. With TPU v6 Trillium already delivering strong inference economics, expect aggressive per-token pricing — potentially undercutting the $5/$25 per million token range that Claude Opus charges.
TPU v7: The Infrastructure Play That Matters Most
The announcement that will get the least press coverage but has the most practical impact: TPU v7 access via Vertex AI. This is the Trillium successor, and it directly affects developer cost structures for running inference at scale.
For teams comparing Google Cloud vs. AWS vs. Azure for AI workloads, TPU v7 pricing and availability could be the deciding factor for 2026-2027 infrastructure decisions.
Android 17 Developer Preview
The Android 17 developer preview will drop alongside the keynote. The AI integration story is the one to watch — Gemini Intelligence (Google's on-device AI layer) is expected to get deeper integration with the OS, including AI-powered widgets, improved Google Assistant, and new developer APIs for building AI features into Android apps.
Project Astra: From Preview to Production
Project Astra is Google's multimodal agent — think real-time video understanding, ambient audio processing, and contextual reasoning across modalities. The I/O announcement is expected to move Astra from preview to production API access, which means developers can start building applications that:
- Process live video feeds in real-time
- Handle multi-turn visual conversations
- Transcribe audio with contextual understanding (not just speech-to-text)
- Combine all of the above in a single API call
This is the capability that could differentiate Gemini from OpenAI and Anthropic in 2026.
Competitive Context: Why This I/O Matters
The AI model landscape in May 2026 is the most competitive it's ever been:
- OpenAI: GPT-5.5 launched in late April, GPT-5.5-Cyber for security use cases
- Anthropic: Claude Opus 4.7 released April 16 with major gains in software engineering (64.3% SWE-bench Pro) and vision accuracy (98.5%)
- DeepSeek: V4 launched April 24 as open-source under MIT license, 1M context window, competitive with frontier models at a fraction of the cost
- Meta: Muse and Spark models expanding the closed-source LLM lineup
- Mistral: Medium 3.5 as an open-weight 128B flagship
Google needs Gemini 3.1 Ultra to be a clear step forward — not just matching the competition but establishing a new benchmark. The Q1 revenue growth suggests they have the momentum to pull it off.
What to Watch For
- Gemini 3.1 Ultra benchmarks — Does it beat Claude Opus 4.7 and GPT-5.5 on SWE-bench and reasoning tasks?
- Pricing — Does Google undercut the competition on per-token costs?
- Astra production access — When can developers actually use it?
- TPU v7 availability — Pricing and timeline for production inference
- Context window — Does Ultra hit the 2-4M token range?
The keynote starts at 10 AM PT on May 19. Livestream at io.google/2026.
We'll have full benchmark analysis and developer guides as soon as Gemini 3.1 Ultra is available for testing.