In 2025, AI image generation costs roughly $0.0006–$0.19 per image depending on tier. Budget API models like Nano Banana 2 Lite run 3 cents per image ($30 per 1,000 images); premium API tiers run $0.02–$0.19 depending on quality settings; self-hosted open-weight models drop to fractions of a cent when you only pay electricity.
Price out your generation by tier — the spread across 2025 options is more than 100x, so the tier you pick matters far more than the exact model within it:
Tier | Cost per image | Notes |
|---|---|---|
Nano Banana 2 Lite | $0.03 | 1K resolution; $30 per 1,000 images; 4-second generation |
Nano Banana 2 | ~$0.06 | Lite is half its cost |
~$0.12 | Roughly 4x Lite's cost at 1,000-image volume | |
Premium API tiers (quality-dependent) | $0.02–$0.19 | Highest quality settings sit at the top of the range |
Self-hosted open-weight (rented GPU) | $0.002–$0.003 | Requires setup and pipeline maintenance |
Self-hosted local (electricity only) | ~$0.0006 | User-reported figures; hardware cost not amortized |
What drives the price differences. Within API models, you pay for output quality and resolution: Nano Banana 2 Lite's 1K resolution is a deliberate design trade-off that buys its 3-cent price, not a quality failure — Google's own benchmarks show it outperforming Nano Banana Pro on text-to-image quality, so the cheapest model is no longer automatically the worst one. Self-hosting undercuts every API price but only pays off at sustained volume — the common break-even heuristic is around 1,000 images per month, below which API pricing wins once you account for setup time and GPU costs.
Budget on cost per usable image, not cost per generation. Raw per-image pricing understates real cost if your rejection rate is high: a $0.03 model that gives you one keeper in ten effectively costs $0.30 per usable image. The standard 2025 optimization is two-tier routing — run all exploration, drafts, and variant testing on a cheap model, then escalate only the selected assets to a premium tier for final quality. In one documented workflow, only the chosen concepts get escalated to Nano Banana Pro, which eliminated internal draft-review costs entirely. Inside invideo, the invideo agent handles this routing across the full image stack — Nano Banana tiers, Recraft, and GPT-Image-2 — so you don't manage per-model billing yourself.
What these prices change in practice. At 3 cents per image, per-generation budgeting stops being a meaningful exercise: a D2C advertiser can test 50 ad concept variants for less than the cost of a coffee instead of budget-capping at 5, find the winner, and cut customer acquisition cost. Filmmakers who previously treated 20 mood-board iterations as a budget conversation now generate them at effectively zero marginal cost. Generating 10 options instead of one becomes the rational default once cost and speed drop out as constraints.
Watch some of these to see what works for you:
if image generations is that cheap and that fast, you don't budget your generations anymore
— invideo's creative team