Nano Banana 2.1 Launches: Specs, Pricing, Benchmarks and Comparisons

Living article · Published October 6, 2026. Launch-day reporting based on official documentation and public benchmark results. Kingy has not run hands-on tests of Nano Banana 2.1. Verified changes will be recorded in the update history below.

Google released Nano Banana 2.1 on October 6, 2026, with a stable API model, lower image-output pricing than Nano Banana 2, and stronger results on public image-generation and editing leaderboards. For developers already using Google’s image models, it is a serious migration candidate. For anyone choosing across providers, the evidence supports comparing it with GPT Image 2.5 rather than assuming Google now leads every task.

The release is confirmed in Google’s October 6 API release notes. The model ID is gemini-nano-banana-2.1. Google also announced that gemini-3.1-flash-image, the stable Nano Banana 2 API model, will shut down on October 29. Existing integrations have a practical reason to evaluate the replacement now.

What changed in Nano Banana 2.1?

Google’s updated product page emphasizes visual design, editing with masks, subject consistency, and more natural images. Those changes target familiar production problems: a character’s face drifting across edits, a product losing its label, or a poster looking attractive while its lettering falls apart.

Our editorial assessment is that preservation deserves as much attention as the first generated image. An advertising team may accept a beautiful draft, then reject the next version because the model changed the packaging. A useful upgrade should survive those revisions. The launch benchmarks give reasons to investigate that possibility; they do not establish the success rate for your own assets.

Nano Banana 2.1 specifications

Source: Google’s model specification. Google reports fixes for tiling artifacts in very wide or tall images at 2K and 4K.

There is a launch-day documentation discrepancy. The DeepMind model card describes a one-million-token context and 64K text output, while the API specification above lists smaller limits. For integration planning, use the documented limits of your endpoint and verify actual behavior before relying on larger figures.

The Google Cloud specification also lists a 131,072-token context and 32,768 maximum output tokens. It documents 15 aspect ratios, including 16:9, 9:16, 21:9, 9:21, 1:8 and 8:1. A “4K” option is a resolution tier, not a promise that every output is a 4096-pixel square.

Costs: image output is cheaper, but the total bill has several parts

The following are Google’s published USD image-output charges. They exclude prompt inputs, text and thinking output, and chargeable search requests.

For 2.1, inputs cost $1.50 per million tokens; text and thinking output cost $7.50. Google lists no Developer API free tier. Search includes 5,000 free requests monthly, shared across Gemini 3.x models, then $14 per 1,000 requests. One generation may cause multiple search queries. See Google’s pricing page.

Calculated from those rates, 1,000 Standard 1K image outputs cost $33.60 before other charges. Treat this as one line of a budget. Track total request cost and the number of usable images you retain. A cheap output that needs three more attempts may be more expensive than a stronger first result.

Google’s benchmarks show better editing and design

The model card reports these vendor-run evaluations. Google uses human side-by-side judgments for preference scores and an automated rater for factuality.

The direction is encouraging across these tasks. Preference scores are not percentages, and Google’s factuality metric is not a guarantee that an infographic is correct. These scores also use a different evaluation pool from Arena, so comparing their numerical levels would be misleading.

Independent rankings: how it compares with other powerful image models

At our October 6 check, Arena’s overall text-to-image leaderboard showed the following snapshot:

*Arena marks these entries preliminary. Nano Banana 2.1’s displayed rank spread is 4–6, so the nearby ordering deserves caution. Rankings, vote counts and model configurations can change.

The single-image editing leaderboard places 2.1 sixth at 1428 ± 6 from 12,985 votes, with a rank spread of 4–7. Sunburst leads at 1524 ± 5 and Flare follows at 1481 ± 5, both preliminary. Nano Banana Pro’s 2K entry is twelfth at 1390 ± 3; Nano Banana 2’s web-search entry is fourteenth at 1387 ± 3.

That supports a narrower conclusion than a universal winner: 2.1 improves Google’s standing, while OpenAI leads these two overall preference tables. Neither table measures your turnaround time, API reliability, total production cost or tolerance for small text errors. Kingy has not independently reproduced these results.

Which model should you evaluate?

Google recommends Nano Banana 2.1 for new projects in place of Nano Banana 2. Its guide continues to position Pro for demanding visual tasks and Lite for speed and cost. For a Google-based workflow, our assessment is to evaluate 2.1 first, then compare Pro on the specific tasks where you already trust it.

For a cross-provider shortlist, include GPT Image 2.5 Sunburst when editing precision matters, and Flare when speed matters. Both list $30 per million image-output tokens and $15 for Batch. That equal token rate does not imply equal per-image bills: image token consumption, quality settings and input charges must be compared separately. OpenAI explicitly warns that its GPT Image 2 calculator does not estimate 2.5 consumption.

A fair evaluation should use identical prompts, the same reference assets, comparable output dimensions and a defined revision sequence. Record latency, total charges and failed requests. Judge the final images without seeing the model names. We would prioritize a product label that must remain unchanged, a four-character scene, a multilingual poster, a targeted masked edit and an infographic with separately checked facts.

Availability and limitations

Google’s distribution list includes Gemini, AI Studio, the Gemini API, Search AI Mode, Ads, Flow and Stitch. The Cloud endpoint is documented separately. These listings do not establish access for every account or region, or identical credits and quotas across products. We have verified documentation, not every product’s account-level rollout.

Google acknowledges small-text blur, imperfect character preservation and errors in masked edits. Review letters, identities and factual labels before using an output. The image-generation guide says generated images include SynthID. Watermarking does not validate the contents of a diagram.

For teams migrating from Nano Banana 2, check hard-coded model IDs, output parsing, resolution settings and retry behavior. Run a small representative evaluation before routing production traffic to 2.1, and use the official deprecation schedule to plan around the October 29 shutdown.

Living update history

October 6, 2026 — Initial publication. Added confirmed release status, API specifications, Standard and Batch pricing, vendor evaluations, independent Arena snapshots, availability and migration guidance. Flagged the context-limit discrepancy. This is source-based reporting; Kingy hands-on Nano Banana 2.1 tests: 0.