Google announced the Gemini 3.5 family and released its first model, Gemini 3.5 Flash, on May 19, casting it as a foundation for agentic work rather than an update to a conventional chatbot. The model immediately became the default in the Gemini app and in AI Mode in Search globally.
The rollout, which coincided with the company’s I/O 2026 developer conference, embeds the speed-oriented Flash variant at the heart of Google’s consumer and enterprise AI infrastructure. It is meant to power coding workflows, long-running tasks and a new category of persistent background agents, not just answer queries.
Google described the model as its strongest yet for complex agentic workflows, coding and long-horizon tasks. A model card published by Google DeepMind lists intended uses in enterprise processes, coding and agentic applications, and discloses safety evaluations conducted under the company’s Frontier Safety Framework.
Developers can access Gemini 3.5 Flash through Google Antigravity, the Gemini API available in Google AI Studio and Android Studio, and enterprise customers can deploy it via the Gemini Enterprise Agent Platform and Gemini Enterprise products. API pricing on the paid tier is $1.50 per million input tokens and $9.00 per million output tokens, which include thinking tokens; batch and flex pricing cut both rates by half. Independent tracker Artificial Analysis measured the model at 284.2 output tokens per second, ranking it second out of 147 models on speed and seventh on an intelligence summary, though it noted that some quality scores are lab-claimed and the model can be verbose.
Alongside the model, Google introduced Gemini Spark, a personal AI agent powered by 3.5 Flash and designed to run continuously in the background. It began rolling out to trusted testers on May 19, with a beta planned the following week for US subscribers of the Google AI Ultra plan.
Google’s performance claims are ambitious. It says 3.5 Flash outperforms the earlier Gemini 3.1 Pro on benchmarks including Terminal-Bench 2.1 (76.2%), GDPval-AA (1656 Elo), MCP Atlas (83.6%) and CharXiv Reasoning (84.2%). It also claims the model is four times faster than other frontier models and often completes agentic tasks at less than half the cost, but the company has not disclosed the detailed methodology or the full set of models compared. Independent benchmark verification is not yet available.
Google expects to launch Gemini 3.5 Pro, a more capable sibling used internally, next month. The move signals rapid family expansion even as the company anchors its immediate agent push on a mid-tier flash variant.
The safety disclosures remain self-reported. No external audit of the agentic risk profile has been conducted, and the hazards introduced by long-running autonomous agents—such as compounding errors or unexpected actions over time—are not yet examined by third parties.
Partner names including Shopify, Macquarie Bank, Salesforce, Ramp, Xero and Databricks appeared in Google’s announcement but have not been separately confirmed. Exact enterprise pricing beyond the public API tiers is unknown, and global availability may still vary by region and account type. For developers and operators considering adoption, the gap between Google’s agentic narrative and independently reproducible evidence is the immediate question.