At Google I/O 2026, Google introduced Gemini 3.5 — its latest family of models built to combine frontier intelligence with action. The series opens with 3.5 Flash, a model Google describes as its strongest agentic and coding release yet, tuned for the complex long-horizon tasks that define real-world utility.
Frontier Performance, Exceptional Speed
3.5 Flash delivers intelligence that rivals large flagship models at Flash-tier speeds. The benchmark slate:
- Terminal-Bench 2.1: 76.2%, outperforming Gemini 3.1 Pro
- GDPval-AA: 1656 Elo
- MCP Atlas: 83.6%
- CharXiv Reasoning: 84.2% for multimodal understanding
Measured on output tokens per second, it is 4 times faster than other frontier models, landing in the top-right quadrant of the Artificial Analysis index — proof, Google argues, that quality and latency are no longer a tradeoff.
Agentic Tasks at Scale
That balance makes 3.5 Flash suited to long-horizon agentic work — tasks that took developers days or auditors weeks, often at less than half the cost of other frontier models. Paired with the updated Antigravity harness, it can deploy collaborative subagents that execute multi-step workflows under supervision: renaming and categorizing unstructured assets, converting legacy codebases to Next.js, even synthesizing the AlphaZero paper and coding a playable game in six hours.
Real-World Impact
Partners already report meaningful deployments:
- Shopify runs parallel subagents for global merchant growth forecasts
- Macquarie Bank pilots faster customer onboarding by reasoning over 100+ page documents
- Salesforce integrates 3.5 Flash into Agentforce for complex, multi-turn tool calling
- Ramp uses it for smarter OCR over complex invoices
- Xero automates multi-week workflows like 1099 tax form gathering
- Databricks runs agents that diagnose issues across massive datasets
Personal AI Agents
3.5 Flash is now the default model for the Gemini app and AI Mode in Search globally, and it powers the new Gemini Spark — a 24/7 personal AI agent that takes action on your behalf under your direction. Spark rolls out to trusted testers first, with a Beta for Google AI Ultra subscribers in the US the following week. On the model front, 3.5 Pro is already in internal use and follows next month.
What This Means
The release shows Google’s 2026 thesis in one model: intelligence matters, but action at speed is what enterprises pay for. With strengthened cyber and CBRN safeguards under the Frontier Safety Framework — including interpretability tools that inspect the model’s inner reasoning — 3.5 Flash sets the pace the rest of the family will follow.