Nano Banana 2.1 on fal bills tokens: $35.294 per million

The fal page for Nano Banana 2.1 (read 2026-10-08) now bills tokens: $1.764 per million input tokens, $8.823 per million output text and thinking tokens, and $35.294 per million output image tokens. It gives about $0.040 per 1K image, $0.059 at 2K and $0.134 at 4K for a brief prompt with medium thinking. Sume's catalog is a flat card per resolution tier, so a Sume call does not move with prompt length or thinking level.

The Sume figures in the repo were set from fal's page on 2026-10-06 at $0.08 for 1K. The current fal page shows a lower, token-based estimate. The pages do not say whether fal changed its price or whether the estimates assume different settings, so the useful fact is where to read what you will be billed.

Side by side

Sume's list rates are in the repo's price tables (0.5K $0.06, 1K $0.08, 2K $0.12, 4K $0.16). The billed column is list times 1.25.

Why the shapes differ

Token billing makes the price depend on the prompt, the references and the thinking level, which fal's page calls out: resolution drives cost, with 4K roughly three times 1K. A flat card is easier to budget: the Sume API docs say cost is the USD amount billed to your wallet, and cost_usd × n is what an endpoint pricing line charges.

The trade is that a short prompt at 1K may cost more on a flat card than on a token meter. For a bulk job that is worth measuring, not guessing.

Read the live number

The endpoint record is the source of truth for Sume. This call returns the billed cost_usd for the row:

import os, requests

r = requests.get(
    "https://api.sume.com/v1/images/models/google/nano-banana-2.1/endpoints",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    timeout=30,
)
r.raise_for_status()
for line in r.json()["endpoints"][0]["pricing"]:
    print(line["billable"], line["unit"], line["cost_usd"])

Note on the 0.5K tier

Sume's contract still lists a 0.5K tier for this row, shown as 512 in the catalog. fal's edit page says the 512px option is discontinued. If you depend on the smallest tier, send one call and check the returned size and the charge before a batch.

The practical difference is predictability. A flat tier lets you multiply n by one number before you send, while a token meter depends on the size and content of the prompt and any references. Sume's endpoint line is the amount billed for one image at the default settings, and usage.cost on each response is the amount actually charged.

If a vendor page and the endpoint disagree, trust the endpoint for what Sume will bill, and re-read both when you plan a large run. The vendor price is a list figure before Sume's multiple, and vendor pages change.