On May 19, 2026, at Google I/O, Google released Gemini 3.5 Flash, the first model in the Gemini 3.5 family, and made it the default in the Gemini app and AI Mode in Search globally. CEO Sundar Pichai framed the launch as the beginning of the "agentic Gemini era," positioning the model around complex coding, multi-step tasks and the ability to coordinate sub-agents.
The release is more than a model update. By embedding the agent-first model as the default on the surfaces used by billions of consumers and thousands of enterprise developers, Google is reshaping how users and companies encounter its AI.
The model succeeds Gemini 3 Flash, which arrived in December 2025, and Gemini 3.1 Flash-Lite from March 2026. Google has pitched 3.5 Flash not as a linear improvement but as a model built from the ground up for agentic workflows, where a system must plan, use tools and act over extended time horizons without constant human prompting.
Gemini 3.5 Flash is available through the Gemini API in Google AI Studio and Android Studio, the Gemini Enterprise Agent Platform, Gemini Enterprise, and the consumer Antigravity service. The Google DeepMind product page labels the model as "Preview," but Google said it is generally available in those named channels. Standard-tier API pricing is $1.50 per million input tokens and $9 per million output tokens; batch pricing halves those rates. In Search, the model is the default only for AI Mode, not for standard web results.
Technically, the model supports a 1-million-token input context window, a 64,000-token output limit and a knowledge cutoff of January 2025. Google describes it as its "strongest agentic and coding model yet," built for function calling, structured output, code execution and Search as a Tool—capabilities that let it orchestrate work across sub-agents and external data.
Pichai also announced Gemini Spark, a 24/7 personal agent powered by Gemini 3.5 Flash. Spark will begin with trusted testers and then launch in beta for Google AI Ultra subscribers in the U.S. the following week. It is not yet broadly available.
Google supplied benchmark scores that it says show 3.5 Flash outperforming the earlier Gemini 3.1 Pro on Terminal-Bench 2.1, GDPval-AA, MCP Atlas and CharXiv Reasoning. Independent benchmarking has not been published. The company also claims the model is four times faster than other frontier models in output tokens per second and costs less than half as much; it did not detail which models were used in the comparison or the measurement methodology.
For safety, Google stated the model was developed under its Frontier Safety Framework, with strengthened safeguards against cyber and CBRN risks, and improved refusal behavior. No independent safety audit has been identified.
The enterprise angle was underlined by a list of names Google put forward—Shopify, Macquarie Bank, Salesforce, Ramp, Xero and Databricks—but none had issued public confirmations of deployment at launch. Pichai offered a hypothetical cost illustration: if large companies shifted 80% of their workloads, totaling about one trillion tokens per day, to the new model, they could save more than $1 billion annually. The figure was presented as illustrative, not guaranteed.
The company disclosed that its 2026 capital expenditures were expected to reach $180–190 billion, reflecting the scale of its AI infrastructure push. Google also said Gemini 3.5 Pro, now in internal use, would launch the following month, and previewed additional agent features for Search and other products without giving specific timelines.
By standardizing on a single agentic model across consumer, developer and enterprise channels, Google is raising the competitive stakes. Rivals will need to match not only model performance but integration depth, speed and cost. Whether real-world reliability, tool-use failure rates and the claimed safety measures can keep pace with that ambition remains an open question.