Google's NB2 Lite undercuts NB1 on price, restricted to 1k resolution
Google is shipping Nano Banana 2 Lite, technically Gemini 3.1 Flash-Lite Image, at $0.034 per 1,000 images with a 4-second generation target, framing it as a low-cost, high-throughput utility for enterprise workflows. According to internal documentation cited in the source, the price undercuts Google's own legacy NB1 ($0.039), runs roughly half the cost of the standard NB2 ($0.067), and a quarter of NB Pro ($0.134). The framing leaves the actual operating envelope narrower: NB2 Lite is restricted to a 1k resolution canvas, the source's internal benchmarks place its text-to-image Elo at 1251, and the deployment is locked to Google's API-only structure, not an open-weights model teams can self-host.
The headline price puts NB2 Lite below Google's own older model on the same line. At $0.034 per 1,000 images, it undercuts the legacy NB1 at $0.039 per 1,000, runs roughly half the cost of the standard NB2 at $0.067, and is about a quarter of the NB Pro tier at $0.134. According to internal notes cited in the source, the model delivers roughly 60-70% of the general capability of NB2 and NB Pro while executing at higher speeds and lower cost. That ratio is the structural argument: a cheaper, faster path that covers most of the heavier models' general utility, with the most expensive tiers reserved for the work the Lite version cannot do.
The 1k-only resolution cap is the obvious constraint. NB2 and NB Pro support 1k, 2k, and 4k output; NB2 Lite is restricted to a 1k canvas. The source does not say what proportion of enterprise image-generation tasks fit a 1k ceiling, and the framing matters here: programmatic A/B testing for ad variations, layout adjustments on localized storefronts, and rapid mockup drafting do not require 4k output, but print, large-format display, and any pipeline that needs downstream upscaling do. The 1k cap effectively partitions the workload between NB2 Lite and the heavier tiers by file size rather than by capability tier.
The Elo numbers narrow the picture further. According to internal benchmarks cited in the source, NB2 Lite posts a text-to-image arena Elo of 1251, against NB1 at 1151 and NB Pro at 1245. That places it ahead of the larger, more expensive NB Pro on a single metric the source flags as "remarkably" edging it out. The source also reports a single-image editing Elo of 1308 and a multiple-image editing score of 1294 for NB2 Lite. None of these are independent evaluations; they are Google's standardized internal benchmarks. The text-to-image win over NB Pro is the headline result, but the model is still scoped to a single resolution and a 60-70% general-capability band, so the practical question is whether that one benchmark tracks the workloads buyers actually have.
The deployment structure is the more durable signal. NB2 Lite is available through Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform (GEAP). The source contrasts this with open-weights models developers can run locally under licenses like Apache 2.0 or modified OpenRAIL, and specifically with Krea's Krea 2 Turbo, a partially open-licensed model that allows open modification and commercial use by small enterprises. The choice the source surfaces is structural: Google's model runs only on Google's metered infrastructure, with usage bound to the pricing terms, while Krea's model can be pulled and modified under its open-license terms. For teams with a self-hosting option already in their stack, the API-only model adds a recurring line item that compounds at the throughput levels the source is targeting.
The production environments the source highlights all favor high-volume, lower-stakes work. Three patterns: world knowledge for drafting contextual scenes and location-specific mockups; character consistency for storyboarding and digital fashion try-ons where object identity needs to hold across sequential generations; and localized typographic rendering for embedding legible copy into ad variants. The unifying feature is speed and iteration, not final-asset quality. The source also notes that conditional image editing may carry marginally higher response times than native generation because of the secondary processing layer required to rewrite existing pixels. The conditional path is exactly the editing use case the source cites, and the latency signal on that path is the place where the 4-second claim is least likely to hold.
The narrow fit is the point: the price is for tasks that stay inside 1k, do not need the heavier models' general capability, and accept metered API access. The source does not test the model under concurrent load, irregular task structures, or distribution shift, and those are the conditions that decide whether a cheap image API stays cheap once the request patterns get messy.