Google DeepMind released its new image model Nano Banana 2.1 on October 6, and the same day flipped it to general availability in the Gemini API under the model ID gemini-nano-banana-2.1. It's an update to Nano Banana 2 (API name gemini-3.1-flash-image), and that older endpoint will be retired on October 29.
According to the official model card, Nano Banana 2.1 sits in the Gemini 3 family, built on Gemini 3.6 Flash as a natively multimodal reasoning model. It accepts text and image input with a context window of up to 1 million tokens, and can output either an image or up to 64,000 tokens of text.
What actually changed
The API changelog lists four improvements: image quality, prompt adherence, character consistency across multi-turn conversations, and text rendering. The new addition is support for ultra-wide aspect ratios — 1:4, 4:1, 1:8 and 8:1 — at 1K, 2K or 4K resolution. Google says speed and cost stay at Flash-tier pricing.
The model card includes two headline comparisons:
- Overall human preference score for text-to-image generation: Nano Banana 2.1's Thinking variant scores 1050 (±14), versus 990 (±7) for Nano Banana 2.
- Factual accuracy on infographics rose from 0.179 to 0.521.
That second number is nearly a threefold gain. Infographics are a notoriously hard case for image models — they have to get the picture right and the numbers and labels right at the same time. Even so, 0.521 still means roughly half of generated infographics contain a factual error, so anything meant for formal use still needs a human check.
What Google admits is still weak
The model card lists its own known limitations: text rendering degrades at small font sizes and in long paragraphs; character consistency between input and output images isn't guaranteed; and spatial reasoning is limited, occasionally mixing up left and right. Knowledge cutoff varies by domain — some areas go up to March 2026, others only to January 2025.
On safety, the model card says Nano Banana 2.1 cleared the child-safety launch bar, that dedicated red-teaming turned up no serious issues, and that frontier safety evaluations didn't trigger any capability level requiring tracking or classified as critical.
Where it's rolling out
Beyond the Gemini app, Google AI Studio and the Gemini API, Nano Banana 2.1 is also wired into Search's AI Mode, Google Ads, the video tool Flow, and the design tool Stitch. The ads and search integrations suggest Google plans to use it directly for bulk-generating commercial creative.
For developers outside the US
The Gemini API isn't directly reachable from mainland China, so teams there typically go through overseas cloud providers or third-party aggregator platforms. With the old endpoint retiring in 23 days, the biggest risk is whether those aggregators remap their models in time; any team still running production traffic on gemini-3.1-flash-image should switch to the new model ID in a test environment this week.
There's a specific wrinkle for Chinese-language output, too. The model card flags poor rendering of small text as a known limitation, and Chinese characters — denser and more stroke-heavy than Latin letters — tend to blur more easily at the same font size. Large headlines should be fine for posters or e-commerce hero images, but small-print captions may still need manual touch-ups afterward. Google hasn't published separate benchmark data for Chinese-text rendering.
The changelog also doesn't mention pricing for the new model, and Google hasn't said whether it will carry over Nano Banana 2's billing rates.
Sources: CocoLoop, Google DeepMind model card, Gemini API changelog. Preference scores and infographic accuracy figures are as published in the model card; the retirement date follows the Gemini API changelog.