Gemini 1.5 Pro-002, Flash-002 and 2.0 Flash update Google AI

Google updated Gemini with 1.5 Pro-002, 1.5 Flash-002, and a Gemini 2.0 Flash preview.

Google rolled out two refreshed Gemini models on September 24, 2024 and followed with Gemini 2.0 Flash Experimental on December 11, 2024. The updates add live multimodal input, native image generation, controllable text-to-speech, Search integration, and new agent and SDK tools.

What changed

Google’s September refresh focused on the 1.5 line, which already used a mixture-of-experts design and a 1 million-token context window. The new Pro-002 and Flash-002 builds were positioned as updated releases rather than a full model family reset.

By December, Google had moved to Gemini 2.0 Flash Experimental, adding features aimed at interactive apps and developer workflows. The model includes a Multimodal Live API for real-time audio and video, native image generation, controllable text-to-speech with watermarking, integrated Google Search, improved agentic behavior, a new Google Gen AI SDK, and Jules, an experimental coding agent for GitHub.

• Sept. 24, 2024: Gemini-1.5-Pro-002 and Gemini-1.5-Flash-002 released.
• Dec. 11, 2024: Gemini 2.0 Flash Experimental announced.
• 2.0 Flash adds live audio/video, image generation, and text-to-speech.
• Google also folded in Search, SDK, and GitHub coding-agent support.

Why it matters

For developers, the shift is less about a benchmark headline and more about a wider API surface. Real-time multimodal input and output means Gemini can be used in voice apps, video-aware assistants, and interactive coding tools without stitching together separate services.

For the market, Google is signaling that Gemini is moving from chat and document tasks into agent-style products that can act, search, and respond across formats. That puts pressure on competing model vendors to match both capability and developer tooling, not just raw model scores.

The bigger question is whether Google can turn these feature drops into sticky app adoption before rivals match the same multimodal and agent features.