Gemini 2.0 Flash
Gemini 2.0 Flash is a fast, low-cost multimodal large language model built by Google DeepMind as the flagship workhorse of the Gemini 2.0 generation, designed for the agentic era with native tool use, a 1 million token context window, and the ability to generate images and audio directly. Google introduced it on December 11, 2024 and stated that the model "even outperforms 1.5 Pro on key benchmarks, at twice the speed" [1]. It launched first as an experimental preview (gemini-2.0-flash-exp) and reached general availability on February 5, 2025, priced at $0.10 per million input tokens and $0.40 per million output tokens [1][3][7].
Gemini 2.0 Flash is distinguished from earlier Gemini models by native multimodal output: a single model accepts text, images, audio, and video as input and can return text, generated images, and synthesized speech without routing to separate systems. It powered Google's early agentic prototypes, including Project Astra and Project Mariner, and was succeeded by the reasoning-native Gemini 2.5 line in March 2025.
Background
Gemini 2.0 Flash succeeded Gemini 1.5 Flash, the fast, cost-efficient tier of the previous generation. Google framed the new model as keeping the speed of 1.5 Flash while pushing quality upward, and it stated that 2.0 Flash "even outperforms 1.5 Pro on key benchmarks, at twice the speed" [1]. The launch was led by Google DeepMind CEO Demis Hassabis and CTO Koray Kavukcuoglu, who positioned the 2.0 family around three themes: multimodality, agentic capabilities, and speed [1][2]. Announcing the release, Google described the goal in terms of agents, writing that "with new advances in multimodality, like native image and audio output, and native tool use, it will enable us to build new AI agents that bring us closer to our vision of a universal assistant" [1].
The December 11 release was deliberately staged. Developers could begin testing an experimental build (gemini-2.0-flash-exp) immediately, while a chat-optimized version appeared in the consumer Gemini product the same day [1][4]. Broader general availability, additional sizes, and the more advanced output modalities were promised for early 2025 [1].
What can Gemini 2.0 Flash do?
Gemini 2.0 Flash accepts text, images, audio, and video as input. Its defining feature relative to earlier Gemini models is native multimodal output: the model can produce text, generate images, and synthesize speech from a single model rather than routing requests to a separate image or text-to-speech system [1][2].
Native multimodal output
At announcement, Google described two new output modes. The first was native image generation, in which the model interleaves generated images with text and supports multi-turn conversational editing, where a user can refine an image across successive prompts [1][5]. The second was steerable, multilingual text-to-speech audio, with Google citing a set of high-quality voices spanning multiple languages [1][2]. Google said all generated images and audio would carry an invisible SynthID watermark for provenance [2].
These output features were restricted to early-access partners at launch rather than being broadly available. Google opened native image generation to general developer experimentation through Google AI Studio and the Gemini API on March 12, 2025 [5][6]. The image, audio, and live-streaming capabilities were described as reaching wider availability "in the coming months" after the February general-availability milestone, which initially shipped with text output [3][7].
Native tool use
Gemini 2.0 Flash can call tools on its own as part of generating a response. Supported tools include Google Search, code execution, and developer-defined functions via function calling [2][3]. Google noted that the model can issue multiple searches in parallel to gather and combine information, which it presented as a building block for agent-style applications [2]. Alongside the model, Google released a Multimodal Live API that handles real-time audio and video streaming input, with support for natural conversational patterns such as interruptions and voice-activity detection [2].
How large is the context window?
Gemini 2.0 Flash provides a context window of 1 million tokens for input [3][7]. That matched the entry tier of the Gemini 1.5 line and was carried over to the budget-oriented Gemini 2.0 Flash-Lite. The larger Gemini 2.0 Pro Experimental, released the same day as the Flash general-availability launch, doubled this to 2 million tokens [3][7].
How does Gemini 2.0 Flash perform on benchmarks?
Google reported that Gemini 2.0 Flash improved on Gemini 1.5 Flash across a range of academic benchmarks and beat the larger Gemini 1.5 Pro on several of them while running roughly twice as fast [1]. The figures below are drawn from Google's published comparison for the experimental release, which compared Gemini 2.0 Flash Experimental with Gemini 1.5 Pro 002 on every benchmark [1].
| Benchmark | Capability | Gemini 2.0 Flash | Gemini 1.5 Pro |
|---|---|---|---|
| MMLU-Pro | General knowledge and reasoning | 76.4% | 75.8% |
| Natural2Code | Code generation | 92.9% | 85.4% |
| MATH | Mathematics | 89.7% | 86.5% |
| MMMU | Multimodal (image) reasoning | 70.7% | 65.9% |
| FACTS Grounding | Factual grounding | 83.6% | 80.0% |
| CoVoST2 (21 languages) | Speech translation (BLEU score) | 39.2 | 40.1 |
| EgoSchema (test) | Video analysis | 71.5% | 71.2% |
Gemini 1.5 Pro 002 still led on two benchmarks in Google's own table: MRCR (1M) long-context understanding (82.6% vs. 69.2%) and CoVoST2 speech translation (40.1 vs. 39.2 BLEU) [1]. Press coverage at launch repeated Google's self-reported figures [8][9].
A separate reasoning-focused variant, Gemini 2.0 Flash Thinking Experimental, was released on December 19, 2024. Built on the 2.0 Flash base, it exposes a summary of its step-by-step reasoning before answering and was widely compared to OpenAI's o1 family of reasoning models [10]. Jeff Dean, then Google DeepMind's chief scientist, described it as "an experimental model that explicitly shows its thoughts" and "trained to use thoughts to strengthen its reasoning" [14].
When was Gemini 2.0 Flash released, and how much does it cost?
During the experimental phase, developers reached Gemini 2.0 Flash through the Gemini API in Google AI Studio and in Vertex AI [1][2]. On the consumer side, a chat-optimized version became selectable in the model drop-down of the Gemini app on desktop and mobile web from December 11, 2024, for both standard and Advanced users, with the mobile app following shortly after [1][4]. On January 30, 2025, Google announced that the Gemini app was switching to Gemini 2.0 Flash as its model, rolling out across web and mobile [17].
At general availability on February 5, 2025, Google described 2.0 Flash as "our highly efficient workhorse model for developers with low latency and enhanced performance" and said it shipped with higher rate limits, stronger performance, and simplified pricing across Google AI Studio, Vertex AI, and the Gemini app [3][7]. The general-availability pricing was set at $0.10 per million input tokens and $0.40 per million output tokens, replacing the earlier split between short and long context tiers [7]. The table below summarizes the launch pricing for the 2.0 Flash tiers.
| Model | Status on Feb 5, 2025 | Input ($/M tokens) | Output ($/M tokens) | Context window |
|---|---|---|---|---|
| Gemini 2.0 Flash | Generally available | $0.10 | $0.40 | 1M tokens |
| Gemini 2.0 Flash-Lite | Public preview | $0.075 | $0.30 | 1M tokens |
| Gemini 2.0 Pro Experimental | Experimental | not priced at launch | not priced at launch | 2M tokens |
What powers Project Astra and Project Mariner?
Google presented Gemini 2.0 Flash as the engine behind several research prototypes shown at the December 2024 launch. Project Astra is a research effort toward a universal AI assistant that uses the model's real-time multimodal understanding through a phone's camera and microphone [1]. Project Mariner was an experimental agent that operated inside the Chrome browser to complete tasks on a user's behalf (Google shut it down on May 4, 2026 [18]); Google reported that it reached a state-of-the-art result of 83.5% on the WebVoyager benchmark, which evaluates agents on end-to-end real-world web tasks [1]. A coding agent prototype, Jules, was also introduced for developers [1].
Variants and successors
The Gemini 2.0 family grew over the following weeks. On February 5, 2025, Google released Gemini 2.0 Flash-Lite into public preview as its most cost-efficient model, offering better quality than 1.5 Flash at the same speed and cost, with the same 1 million token context window [3][7]. Flash-Lite then reached general availability for production use in Google AI Studio and Vertex AI on February 25, 2025 [15]. The February 5 announcement also brought Gemini 2.0 Pro Experimental, aimed at coding and complex prompts, with a 2 million token context window and tool calling for Google Search and code execution [3][7].
The next generation arrived with Gemini 2.5. Google released Gemini 2.5 Pro Experimental on March 25, 2025, a model that built reasoning into the architecture while retaining native multimodality [11]. Google shut down the gemini-2.0-flash and gemini-2.0-flash-lite API identifiers on June 1, 2026; as of November 2026, its deprecations page lists gemini-3.6-flash and gemini-3.1-flash-lite as the recommended replacements [13].
Continue reading
- Gemini (language model)Gemini is a family of natively multimodal large language models developed by Google DeepMind, first announced on December 6, 2023, that can reason across text, images, audio, video, and code within a single…Mentioned 52 times above
- GoogleGoogle LLC is a United States technology company and the largest operating business within Alphabet Inc. Its activities include internet search, online advertising, video, mobile and web platforms, consumer…Mentioned 37 times above
- Multimodal ModelA multimodal model is a machine learning model, or a model-based system, that processes, relates, or produces information across more than one kind of data. Each kind is called a modality.Mentioned 7 times above
- Context windowA context window is the finite token sequence that a language model can process for one invocation.Mentioned 6 times above
- Google DeepMindGoogle DeepMind is an artificial intelligence research and model-development organization at Google.Mentioned 3 times above
- Tool useTool use in artificial intelligence is the ability of a model-based system to request, coordinate, and use capabilities outside the model's ordinary token-generation process.Mentioned 3 times above
More to read
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18Google. "Introducing Gemini 2.0: our new AI model for the agentic era." The Keyword (Google blog), December 11, 2024. blog.google
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Google for Developers. "The next chapter of the Gemini era for developers." Google Developers Blog, December 11, 2024. developers.googleblog.com
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Google. "Gemini 2.0 is now available to everyone." The Keyword (Google blog), February 5, 2025. blog.google
- ^1 ^2Abner Li. "Gemini 2.0 Flash announced with agentic focus, coming to app." 9to5Google, December 11, 2024. 9to5google.com
- ^1 ^2Google for Developers. "Experiment with Gemini 2.0 Flash native image generation." Google Developers Blog, March 12, 2025. developers.googleblog.com
- ^Abner Li. "You can now test Gemini 2.0 Flash's native image output." 9to5Google, March 12, 2025. 9to5google.com
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Google for Developers. "Gemini 2.0: Flash, Flash-Lite and Pro." Google Developers Blog, February 5, 2025. developers.googleblog.com
- ^DeepLearning.AI. "Google Introduces Gemini 2.0 Flash, a Faster, More Capable AI Model." The Batch. deeplearning.ai
- ^Analytics Vidhya. "Gemini 2.0 by Google is Here: Faster and Smarter than Ever Before." December 2024. analyticsvidhya.com
- ^Kyle Wiggers. "Google releases its own 'reasoning' AI model." TechCrunch, December 19, 2024. techcrunch.com
- ^Wikipedia contributors. "Gemini (language model)." Wikipedia. en.wikipedia.org
- Helicone. "Gemini 2.0 Flash Explained: Building More Reliable Applications." helicone.ai
- ^Google. "Gemini API deprecations." Gemini API documentation. ai.google.dev
- ^Jeff Dean. "Introducing Gemini 2.0 Flash Thinking, an experimental model that explicitly shows its thoughts." X (formerly Twitter), December 19, 2024. x.com
- ^Google for Developers. "Start building with Gemini 2.0 Flash and Flash-Lite." Google Developers Blog, February 25, 2025. developers.googleblog.com
- ^Google. "Gemini logo (Google Image Library)". Google Image Library publication permission. Original official raster, unchanged.
- ^Patrick Kane. "Try Gemini 2.0 Flash in the Gemini app." The Keyword (Google blog), January 30, 2025. blog.google
- ^Android Authority. "Google quietly kills Project Mariner as the AI agent race shifts gears." May 6, 2026. androidauthority.com
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
8 revisions · v8 · 1,916 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check of the full article against its cited sources (62 claims, 19 sources fetched); 7 material + 7 minor fixed
Cite this page: AI Wiki. "Gemini 2.0 Flash." aiwiki.ai, updated 24 Jun 2026, fact-checked 2 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/gemini_2_0_flash



