Google Slashes AI Image Generation Costs by 50% with Nano Banana 2.1 and EmbeddingGemma 2 Launch

Google released two AI models that signal a strategic push toward cost-efficient, high-quality image generation and on-device multimodal capabilities. Nano Banana 2.1 and EmbeddingGemma 2 represent significant upgrades in both performance and accessibility for developers and enterprises.1

1

Nano Banana 2.1 Cuts Image Generation Costs in Half

Google's Nano Banana 2.1, built on Gemini 3.6 Flash, delivers substantial cost reductions for AI image generation through the Gemini API. Developer pricing drops to $0.0336 for a standard 1K image, down from $0.067 with Nano Banana 2. A 4K image now costs $0.0756 versus $0.151 previously. Batch processing offers an additional 50% discount for high-volume generation, making a thousand images cost approximately $33.60 through the API.1

1

3

The model handles 1K, 2K, and 4K outputs while supporting unusually wide and vertical image formats as wide as 8:1. It processes up to 14 reference images at once, tracking up to four characters and ten objects across multiple editing turns. This capability proves crucial for maintaining character consistency in sequential edits.2

2

5

Enhanced Visual Quality and Editing Precision

Nano Banana 2.1 scored 1,050 ELO Points in text-to-image tests, outperforming Nano Banana 2's 990 points and Nano Banana Pro's 935 points. The ELO system provides a qualitative measure without a ceiling, indicating clear performance gains.1

1

Google points to three major upgrades: improved visual design, sharper mask-based editing where marked regions change while preserving surrounding details, and better subject consistency. The model resolves tiling artifacts on wide and panoramic aspect ratios at 2K and 4K resolutions. Better text rendering now supports posters, labels, infographics, and advertisements more effectively.2

2

3

Developers can set thinking levels from minimal to high, controlling how long the model processes before generating images. Search grounding allows Nano Banana 2.1 to use information from Google Search and Google Image Search before drawing, improving accuracy for real-world subjects and events.1

1

3

EmbeddingGemma 2 Brings Multimodal Search to Local Devices

Google launched EmbeddingGemma 2, a lightweight multimodal embedding model with 740 million parameters built on the Gemma 4 architecture. Released under the commercially permissive Apache 2.0 license, it processes text, code, images, audio, and video in a unified embedding space optimized for on-device inference.2

2

4

The model requires just 270 million parameters for text-only workloads, with optional vision and audio encoders adding 170 million and 300 million parameters respectively. On a Google Pixel 11 Pro, the text-only version uses approximately 191MB of active RAM, while the full multimodal model requires about 567MB with quantization.2

2

4

EmbeddingGemma 2 features an 8,000-token context window, four times larger than its predecessor EmbeddingGemma, which has been downloaded more than 20 million times. This expanded window processes up to 5.5 minutes of audio, 29 images, 58 video frames, or combinations of these inputs on local hardware.2

2

4

Related Stories

Google launches Nano Banana 2 Lite, its fastest and cheapest AI image generator at 4 seconds

Google's Nano Banana 2 and Gemini 3 Pro Set to Revolutionize AI Image Generation and Coding

Google releases Nano Banana 2 AI image model with Pro quality at Flash speed

Google's Gemini 2.5 Flash Image Model Goes Mainstream with Enhanced Features

Performance Gains and Storage Efficiency

EmbeddingGemma 2 scored 78.68 on the Massive Text Embedding Benchmark Code test, a 9.92-point improvement from EmbeddingGemma's 68.76 score. Google claims leading scores among multimodal embedding models with fewer than one billion parameters across benchmarks including MTEB Code and MAEB.2

2

4

The model uses Matryoshka Representation Learning (MRL), allowing developers to reduce its 768-dimensional output vectors to 512, 256, or 128 dimensions. This reduces storage requirements and memory needs for local vector databases by up to six times, making it practical for resource-constrained environments.2

2

4

Broad Platform Availability and Integration

Nano Banana 2.1 is now accessible in the Gemini app, Google Search AI Mode, Google Ads, AI Studio, Flow, and Stitch. The update includes C2PA Content Credentials for digital origin tracking across supported workflows.2

2

3

EmbeddingGemma 2 weights are available through Hugging Face and Kaggle, with Gemini Enterprise Agent Platform Model Garden availability coming soon. The model deploys via cross-platform apps with Google AI Edge MediaPipe, LiteRT, transformers.js, and WebGPU. It's compatible with MLX, vLLM, llama.cpp, SGLang, Ollama, and LMStudio.2

2

4

When paired with generative models such as Gemma 4, EmbeddingGemma 2 enables on-device RAG pipelines. The model suits local codebase indexing, semantic code search, and coding agent retrieval. Users can find specific video clips from voice memos or search through audio recordings using text queries.2

2