GPT Image 2.5 medium is 6x cheaper than Nano Banana 2 1K on Sume

On Sume, a GPT Image 2.5 image at medium quality costs $0.0165 at 1024x1024, and a Nano Banana 2 image at its 1K tier costs $0.1000, so Nano Banana 2 is 6.1 times as expensive. The ratio holds at the larger tiers: 8.3 times at 2K against 2560x1440 and 6.2 times at 4K against 3840x2160.

This is a price comparison and not a quality ranking. The two models take different parameters, return different looks, and you should run your own prompts through both before you move a workload. What the price gap does tell you is how much a switch can save, or how much a keep-it-as-is decision costs.

The ratio at each size

Each row pairs the Nano Banana 2 resolution tier with the closest GPT Image 2.5 pixel size. Nano Banana 2 prices by tier, so its tier price is the same whatever the aspect ratio. GPT Image 2.5 prices by output tokens, so the size changes the figure.

Why the gap is this wide

The Fal list rate for Nano Banana 2 is $0.08 per image at 1K and rises to $0.12 at 2K and $0.16 at 4K. Sume's price is that list rate times 1.25, which gives the $0.1000, $0.1500 and $0.2000 above. GPT Image 2.5 is billed per token: $30 per million output image tokens, and at medium quality a 1024x1024 image comes to 439 output tokens by Sume's estimator, which is the $0.0132 provider cost that becomes $0.0165 after the 1.25 factor.

Because the GPT price follows tokens, quality is the lever. At the same 1024x1024 size low is $0.0074, high is $0.0659 and max is $0.2635. Only xhigh and max go above Nano Banana 2's 1K price. At high, GPT Image 2.5 is still 34% below it.

What the cheaper row does not do for you

Check these before you reroute, since a price difference is easy to eat in rework.

• Aspect ratios. Nano Banana 2 lists 15 ratios in Sume's catalog, including 21:9 and 1:8, and takes them as aspect_ratio. GPT Image 2.5 takes pixel sizes with a 3:1 limit.
• Resolution tiers. Nano Banana 2 offers 0.5K, 1K, 2K and 4K as named tiers. GPT Image 2.5 offers any size within 655,360 to 8,294,400 pixels, so you choose the exact pixels.
• Quality control. Only GPT Image 2.5 has the six-step quality setting, and Sume's docs list background for GPT Image 2.5 alone.
• Look. Both families return up to four images per call, but their outputs differ, and only your own review can say which matches your brand.

Reproduce the ratios

Use the script to rerun the three pairings when either price changes. Take the figures from GET /v1/images/models or from usage.cost in a test response.

SUME_USD = {
    "gpt-image-2.5 medium 1024x1024": 0.0165,
    "gpt-image-2.5 medium 2560x1440": 0.0180,
    "gpt-image-2.5 medium 3840x2160": 0.0325,
    "nano-banana-2 1K": 0.10,
    "nano-banana-2 2K": 0.15,
    "nano-banana-2 4K": 0.20,
}

pairs = [
    ("nano-banana-2 1K", "gpt-image-2.5 medium 1024x1024"),
    ("nano-banana-2 2K", "gpt-image-2.5 medium 2560x1440"),
    ("nano-banana-2 4K", "gpt-image-2.5 medium 3840x2160"),
]
for dear, cheap in pairs:
    print(f"{dear} costs {SUME_USD[dear] / SUME_USD[cheap]:.1f}x {cheap}")

How to decide

Run a small blind test first. Take twenty prompts that look like your real workload, render each on both models at the pairing you would actually ship, and have a reviewer pick the better image without knowing which is which. If the reviewer cannot tell them apart often enough, the 6x gap is money you can keep.

If the reviewer does prefer Nano Banana 2, the gap is the price of that preference, and you can put a number on it. At 10,000 images a month the difference between the two 1K and 1024 rows is $835.00, so a preference has to be worth that to justify the spend. If it is only worth that on some prompts, route those prompts to Nano Banana 2 and send the rest to GPT Image 2.5 at medium.

Both models are reachable through the same endpoint, so routing is a one-field change in the request body, not a second integration.