Nano Banana 2.1: Flash Image Stack Refresh, Half the 1K Price

Nano Banana 2.1 is Google's Flash-tier image stack refresh: GA on October 6, 2026 as gemini-nano-banana-2.1, based on Gemini 3.6 Flash. It replaces deprecated Nano Banana 2 (gemini-3.1-flash-image) while Nano Banana Pro (gemini-3-pro-image) stays the expensive Pro tier.

Lead with Google's own human-preference Elo and Infographic Factuality (self-reported), then verified Gemini API per-image list prices. Biggest card jumps: multi-character consistency, infographic factuality, and mask-based editing. Pricing story: about half of Nano Banana 2 at 1K/2K; 4K is $0.113 because the 4K token count rose to 3,780.

At a Glance

• Catalog / API id: gemini-nano-banana-2.1
• Organization: Google DeepMind (Gemini 3 family)
• Status: Stable / GA October 6, 2026 (no preview)
• Base: Gemini 3.6 Flash
• Surfaces: Gemini API + Gemini Enterprise Agent Platform (also listed on Gemini App, AI Studio, Search AI Mode, Ads, Flow, Stitch)
• Inputs / outputs: text, image, video, PDF in; image + text out
• Tokens (API): 131,072 in / 32,768 out
• Resolutions: 1K / 2K / 4K (default 1K)
• References: up to 14 images (4 characters / 10 objects)
• Thinking: minimal / medium (default) / high
• Grounding: Google Web + Image Search
• Pricing (Standard image): $0.0336 / $0.0504 / $0.113 per 1K / 2K / 4K
• Self-reported Thinking: Preference Elo 1050 · Factuality 0.521 · Multi-Char 1106
• Predecessor: Nano Banana 2 deprecated; no official shutdown date

What's New

Google positions 2.1 as an update to Nano Banana 2 with Flash-level speed and cost, plus gains in visual quality, prompt adherence, multi-turn character consistency, text rendering, and wide aspect ratios. It is the more efficient counterpart to Nano Banana Pro, not a new Pro tier.

Capabilities worth quoting

• Thinking levels: minimal, medium (default), high
• Up to 14 reference images (4 characters / 10 objects)
• Aspect-ratio tiling fix for 1:4, 4:1, 1:8, and 8:1 at 2K and 4K
• Google Web and Image Search grounding; Batch supported
• Not supported on the Gemini API model page: function calling, structured outputs, caching, Maps grounding, Live API, audio generation

Platform nuance: Cloud Enterprise docs list implicit context caching as supported while the Gemini API model page lists caching as not supported. For Gemini API product facts and pricing, use the API page.

Benchmarks

All scores below are self-reported from the DeepMind Nano Banana 2.1 model card (October 2026). Methodology: side-by-side human eval into Elo on preference / design / editing axes; single-sided AutoRater on factuality. Columns: Nano Banana 2.1 Thinking / No Thinking, Nano Banana 2 Thinking, Nano Banana Pro. Not LLM Stats verified. Catalog benchmarks for this model are empty; these numbers live in the article and charts only.

Text-to-image (self-reported)

With Thinking on, Overall Preference Elo moves from NB2's 990 to 1050, and lands ahead of Pro's 935 on Google's own table. Infographic Factuality is the sharpest relative jump: 0.521 vs NB2 0.179 and Pro 0.265.

Editing (self-reported)

Multi-Character Consistency 1106 and Mask / Ink-Based Editing 1049 are the clearest editing headlines versus both NB2 Thinking and Pro. Prefer Thinking columns when citing a single number; No Thinking trails but still beats NB2 on every row above.

Pricing

First-party Gemini API Standard rates for gemini-nano-banana-2.1. Free tier is not available. Batch is 50% of Standard.

Input is $1.50 / 1M; text + thinking output is $7.50 / 1M. Grounding (paid): 5,000 free search requests/month shared across Gemini 3.x, then $14 per 1,000 requests. Source: Gemini API pricing.

When to Use It

• Good fit:Flash-cost image generation and editing through the Gemini API when you want Google's self-reported preference / consistency gains over Nano Banana 2 without Pro sticker prices.
• Migrate from NB2: deprecated gemini-3.1-flash-image → gemini-nano-banana-2.1. No official shutdown date; still prefer 2.1 for new work.
• Prefer Nano Banana Prowhen you explicitly want the Pro tier and are willing to pay $120 / 1M image tokens ($0.134 at 1K/2K, $0.24 at 4K). On Google's own Thinking Elo table, 2.1 leads Pro on the preference and editing rows above.
• Turn Thinking up for the card peaks (Overall Preference, Multi-Char, Factuality). Default is medium.
• Use 2K/4K panoramic ratios (1:4, 4:1, 1:8, 8:1) when NB2 tiling was a problem; Google lists those artifacts fixed at 2K/4K.
• Budget 4K as $0.113, not "half of NB2 4K." Batch if latency allows (half Standard).

Caveats

• Self-reported only. Elo and factuality are Google DeepMind model-card numbers, not LLM Stats / ZeroEval verified. Catalog benchmarks array is empty for this model.
• Token limit conflict. Prefer API/Cloud 131,072 / 32,768. Model-card prose mentioning 1M / 64K is inconsistent; do not lead with it.
• 4K token revision. 4K rose to 3,780 tokens (NB2 was 2,520). Cite verified $0.113, not a vague half-price claim.
• Model-card limitations. Poor small-text rendering (often blurry at 1K); imperfect character consistency; partial instruction following / ink persistence; rare persistent subject pose; spatial left/right confusion; limited advanced world knowledge / 3D / factuality.
• No invented peers. Official pages only head-to-head vs Nano Banana 2 and Nano Banana Pro. Do not invent Midjourney / Flux / GPT Image comparisons from this drop.
• NB2 shutdown. Deprecated with no date announced. Ignore third-party Oct 29 shutdown claims unless Google updates deprecations.

Outlook

This post ages when LLM Stats verifies preference or editing evals, when Google sets a Nano Banana 2 shutdown date, or when API token limits and model-card prose are reconciled. Through-line: Flash image refresh at gemini-nano-banana-2.1, self-reported Thinking Elo ahead of NB2 and Pro on Google's table, and verified Standard per-image prices of $0.0336 / $0.0504 / $0.113.

Sources: Gemini API model page, DeepMind model card, Gemini API pricing, Gemini API changelog.