Gemini 2.0 Explained

What it is

Gemini 2.0 is Google's model family built for what the company called the agentic era: models designed to understand more about the world, think multiple steps ahead, and take actions on the user's behalf. The headline advances over the 1.x generation were native image and audio output and native tool use, including Google Search, code execution, and user-defined functions.

Google announced the family on December 11, 2024, with an experimental Gemini 2.0 Flash as the first release. The framing was deliberate. Where earlier launches had emphasized benchmark scores, this one emphasized agency: the models were supposed to do things, not just answer questions. General availability followed on February 5, 2025, adding 2.0 Flash-Lite and the experimental 2.0 Pro. The models were built by Google DeepMind and announced by CEO Sundar Pichai and DeepMind CEO Demis Hassabis.

How it works

Gemini 2.0 is a natively multimodal model with agentic capabilities built in: multimodal reasoning, long context, complex instruction following, and compositional function calling. Google disclosed that training and inference ran entirely on its sixth-generation Trillium TPUs, a detail that underscored how much of Google's AI advantage is hardware. Deeper architecture details were not published.

The family powered Google's agent prototypes of the period. Project Astra was the multimodal assistant that could see through a phone camera and hold a conversation about the physical world. Project Mariner operated a web browser on the user's behalf and scored 83.5 percent on the WebVoyager benchmark of real-world web tasks. Jules was the coding agent. A Multimodal Live API supported real-time audio and video streaming conversations, the infrastructure for voice-first applications.

It accepts text, image, video, and audio input, and produces text, image, and steerable text-to-speech audio output natively rather than through bolt-on tools. Native output mattered because it preserved information across modalities instead of losing it at each conversion step.

Key details

Context windows were large: one million tokens for 2.0 Flash and Flash-Lite, two million for 2.0 Pro Experimental. At the time, the two-million-token option was among the largest context windows available anywhere, enough to hold entire codebases or book-length documents in a single prompt.

The knowledge cutoff was June 2024 according to Google's own model card, though some third-party trackers listed August; the model card is the primary source and the safer citation. Google's official benchmarks for 2.0 Flash included 75.8 percent on MMLU-Pro, 59.1 percent on GPQA Diamond, and 86.5 percent on MATH. Google reported that 2.0 Flash outperformed 1.5 Pro on key benchmarks at twice the speed, positioning it as the rare upgrade that was both better and cheaper. These are Google's figures.

Pricing and access

The API pricing Google published was $0.10 per million input tokens and $0.40 per million output for 2.0 Flash, and $0.075 and $0.30 for Flash-Lite. That aggressive pricing was part of the strategy: make the fast model nearly free and win on volume. Google AI Studio offered a free tier with rate limits.

In the Gemini app, 2.0 Flash was available to all users including the free tier, while 2.0 Pro Experimental required the Advanced subscription at around $19.99 per month through Google One AI Premium. Pricing changes over time, so verify current rates before relying on these. One caveat from the research: the 2.0 Flash and Flash-Lite shutdown was reported for June 2026, so check availability before building on these models.

Limitations

The knowledge cutoff of June 2024 meant the model needed search grounding for anything newer, which its native tool use was designed to provide. Google's own model card listed known weaknesses: hallucinations and limited ability in causal understanding, complex logical deduction, and counterfactual reasoning. Those are the tasks where the model sounds confident while being wrong, and no amount of context window fixes them.

Native image generation was initially limited to early-access partners, and in March 2025 there were reports of the experimental image model being used to strip watermarks from copyrighted images. That reporting was secondary, not Google-admitted, but it illustrated the dual-use problem that follows every capable image model. The 2.0 Pro model stayed experimental at launch and never got a full general release in the family. Gemini 2.0 now reads as a transitional generation: the moment Google stopped selling models as answer engines and started selling them as agents.

The agent prototypes

Gemini 2.0's real significance was as the engine for Google's agent experiments, which were the most ambitious in the industry at the time. Project Astra demonstrated a multimodal assistant that could see through a phone camera, remember where things were, and hold a natural conversation about the physical world. The demos were carefully staged, but the direction was clear: the assistant was supposed to perceive, not just respond.

Project Mariner took the browser-operating approach, navigating real websites to complete tasks. Its 83.5 percent score on the WebVoyager benchmark of real-world web tasks was, by Google's reporting, a strong result for an autonomous web agent in early 2025. Jules applied the same agentic thinking to software development, working through coding tasks asynchronously. Together the three projects mapped the space the whole industry would spend the next two years exploring: see, browse, and code.

The Multimodal Live API deserved attention as infrastructure. Real-time audio and video streaming conversations required low-latency model serving that most providers could not offer, and Google's TPU fleet was part of why it could. Voice-first applications built on the Live API previewed the conversational interfaces that later became mainstream.

The family was transitional by design. With a shutdown reported for June 2026 and the 2.5 generation taking over, Gemini 2.0 reads in retrospect as the moment Google committed to agency as the product direction. The benchmarks mattered less than the bet: that the next era of models would be judged on what they could do, not just what they could say.