Can a cheap AI image model outperform a premium one on quality benchmarks?

Yes — and it has already happened. According to Google's own benchmarks, Nano Banana 2 Lite outperforms Nano Banana Pro on text-to-image quality, despite costing 3 cents per image, roughly one-quarter of Pro's price. Cheap no longer means worst; it means deliberate trade-offs, so verify benchmarks against your own prompts before committing.

Start with the documented case: Google's own benchmarks show Nano Banana 2 Lite beating Nano Banana Pro on text-to-image quality. Lite runs at 3 cents per image, generates in about 4 seconds, is 2.5x faster than Nano Banana 2, and works out four times cheaper than Nano Banana Pro at 1,000-image volume — around $30 for 1,000 images. The price gap no longer maps to a quality gap on the core generation benchmark.

Why a cheap model can win a quality benchmark. Budget tiers are built with deliberate trade-offs, not across-the-board downgrades. Nano Banana 2 Lite's 1K output resolution is a design decision — the model spends its capacity on prompt adherence and image quality, which is exactly what text-to-image benchmarks measure. Premium tiers like Nano Banana Pro keep their advantage on resolution and technically demanding shots, dimensions many headline benchmarks don't weight heavily. So "outperforms on quality benchmarks" is true and still compatible with the premium model being the right choice for specific final assets.

Treat benchmark rankings as a starting point, not a verdict. Public rankings can diverge from your real workload depending on prompt type, and practitioners increasingly flag benchmark saturation — once most models score near the top, the benchmark stops discriminating between them. Before committing to a model, run a small side-by-side evaluation on your own prompts: generate the same 10–20 briefs on both tiers and judge outputs blind. At 3-cent pricing, that evaluation costs under a dollar.

What to do with the result. When the cheap tier matches or beats the premium tier on your prompts, route by task instead of defaulting to premium: use the budget model for all high-volume exploration — one documented workflow tests 50 ad concept variants instead of 5 to find a winner and cut customer acquisition cost, for less than the cost of a coffee — and escalate only the chosen assets to Nano Banana Pro for final production quality. Filmmakers apply the same split to pre-production: iterating on 20 mood boards used to be a budget conversation and is now effectively free at Lite pricing. Inside invideo, the full image-model stack — Nano Banana tiers, Recraft, GPT-Image-2 — is available in one place, and the invideo agent routes each task to the appropriate tier so you don't manage the split manually.

The practical answer: yes, a cheap model can top a premium one on quality benchmarks — but confirm it on your own prompt set, and keep the premium tier in the loop for the shots that genuinely need it.

Watch some of these to see what works for you:

the cheapest model is no longer the worst model

— invideo's creative team