Gemini 3.5 Flash Targets Agentic Workflows, Boosts Google Products

Google DeepMind has introduced Gemini 3.5 Flash, the latest addition to its Gemini family of models. This model is engineered for high-volume, long-horizon agentic applications. It is now the default model within the Gemini app and Google Search AI Overviews.

Gemini 3.5 Flash exhibits a 2x speed improvement over Gemini 1.5 Pro. This increased processing speed is coupled with a context window of 262,144 tokens. This large context window supports processing extensive documents, codebases, and videos. The model’s enhanced capabilities aim to facilitate complex multi-step reasoning.

Evaluations indicate Gemini 3.5 Flash performs strongly on agentic benchmarks. These benchmarks include Fido, an internal Google DeepMind multi-step reasoning benchmark, and Agent Atlas, a composite benchmark assessing tool use and multi-step reasoning. The model demonstrates a 23.9% improvement on Agent Atlas compared to Gemini 1.5 Pro. It also achieves a 100% success rate on the Fido benchmark. For comparison, Gemini 1.5 Pro achieved a 93% success rate on Fido, and GPT-4o achieved 89%. Additionally, Gemini 3.5 Flash secured top-tier results in a needle-in-a-haystack evaluation, retrieving embedded information from documents up to 256,000 tokens with 100% recall.

Gemini 3.5 Flash is now accessible via the Google AI Studio and Vertex AI platforms. Its integration into the Gemini app allows for advanced multi-step reasoning features. This enables the model to process nuanced prompts and access real-time information through Google Search. Developers can also use the model to build applications such as summarisation tools and customer service agents. How will this focus on agentic workflows redefine the development landscape for AI applications?