Google ships Gemini 2.0 Flash and agent prototypes

Google released Gemini 2.0 Flash in experimental form and used the announcement to introduce a set of agent prototypes it framed as the start of an “agentic era” for the model family. Flash could take and generate images, video and audio natively alongside text, supported multilingual text-to-speech, and could call tools including Google Search and a code execution environment directly. Google said the experimental version outperformed the previous generation’s larger Gemini 1.5 Pro model on key benchmarks while running roughly twice as fast; the model was available immediately through Google AI Studio and Vertex AI, with general availability and additional model sizes promised for January 2025.

Alongside Flash, Google introduced several agent prototypes built on the new model. Project Mariner, an experimental Chrome extension, let the model view and act on a browser tab to complete web tasks. Deep Research used the model’s reasoning and long-context ability to plan and carry out multi-step web research, compiling findings into reports for Gemini Advanced subscribers. Jules, a third prototype, was pitched as an experimental coding agent that could take on GitHub issues under a developer’s supervision, working asynchronously rather than as a chat-based assistant.

Google positioned the release as a direct response to the reasoning and agent work competitors had shown through 2024 — OpenAI’s o1 reasoning models and various browser- and computer-use agents from OpenAI and Anthropic among them — combining a faster core model with a family of task-specific agents rather than a single flagship release. Most of the agent prototypes launched to small groups of trusted testers rather than to the public, continuing a pattern in which Google previewed agentic capability well ahead of general availability, with broader rollout following through 2025.