Google quietly rolled out Nano Banana 2.1, its newest image generation and editing model built on Gemini 3.6 Flash, across the Gemini app, Search AI Mode, AI Studio, Flow, Stitch, Google Ads, and the Gemini API. The API model ID is gemini-nano-banana-2.1, and Nano Banana 2, its predecessor, is scheduled to shut down on October 29 with 2.1 as the designated replacement. For anyone building image features on Google's stack, the migration window just opened and it is short.
The headline change is economic. Standard API image-output pricing dropped to $30 per million image tokens, so a 1K image now costs about $0.0336 versus $0.067 before; 2K runs $0.0504 and 4K $0.0756. Batch jobs take another fifty percent off, landing 1K images at roughly $0.0168. If you generate marketing assets, product illustrations, or campaign variants at volume, Google just halved one of the more stubborn line items in your pipeline budget.
Capability-wise, the release pushes image editing toward something a production workflow can actually depend on. Outputs reach 4K, including ultra-wide 8:1 aspect ratios. The model ingests up to 14 reference images at once and tracks as many as four characters and ten objects across multi-turn edits, which attacks the identity drift that has made iterative generation so fiddly. Mask-based editing alters a selected region instead of regenerating the whole frame. Developers can also dial the inference thinking level from minimal to high and toggle grounding through Google Search and Google Image Search, so renders can reflect verifiable real-world details instead of hallucinating them.
Treat Google's own numbers with a practitioner's skepticism. Its cited preference test gives 2.1 a 1,050 Elo against 990 for its predecessor, and its infographic factuality score of 0.521 roughly triples the older model's 0.179, but no independent external evaluations were disclosed at launch. Early third-party impressions are mixed, with some testers still reaching for the pricier Pro tier on demanding work. One community roundup put it plainly: test your own prompts before migrating production traffic. Google's claim of outperforming previous models across the board is a marketing line until you verify it on your workload.
The durable signal is the pattern. Google is turning image generation into cheap infrastructure, the same playbook it ran with embeddings. C2PA Content Credentials ship in the box for provenance, a quiet admission that synthetic imagery at this volume needs auditability. Between the price cut, mask editing, and search grounding, 2.1 is less a research milestone than a workbench upgrade: the image step in an agent pipeline stops being the expensive, flaky part and becomes boring. That is exactly when a capability starts showing up everywhere.
