Why CometAPI's image price cards need different budget units

The most tempting image-model comparison is also the least useful one for a budget: putting CometAPI’s Nano Banana 2 dollar-per-image figure beside GPT-Image-2.5’s dollar-per-million-token figure and choosing the smaller number. Those numbers have different denominators. Before picking a model, decide whether the budget needs a quoted price per generated asset or an estimate from measured token usage.

In its September 21 comparison, CometAPI lists Nano Banana 2 at $0.0360 per image at 0.5K, $0.0536 at 1K, $0.0808 at 2K, and $0.1208 at 4K. These are published CometAPI resolution tiers. They let a planner start with the intended output tier and count of generations. They do not, by themselves, tell you how many generations an approval process will need or the total cost of a finished campaign.

The same comparison lists CometAPI GPT-Image-2.5 at $4 per million text-input tokens and $24 per million image-output tokens. It says both Flare and Sunburst show those rates for the comparable dimensions, but directs readers to the current provider quote for image input. Its GPT-Image price is also quality and token dependent. The published rate is therefore a multiplier, not a per-image invoice. The article gives no fixed image-token count for a particular image, quality setting, or edit sequence.

Build two budget columns

For Nano Banana 2, label the column generated images by resolution. Multiply the number of planned generations in each tier by that tier’s quoted per-image price, then add the tier subtotals. Keep revisions in the generation count: one accepted asset may involve several attempts. That is a planning method, not a prediction of how many attempts your team will make.

For GPT-Image-2.5, label the column tokens consumed at the selected quality. At the rates in the article, the text-input component is text-input tokens divided by one million, multiplied by $4; the image-output component is output tokens divided by one million, multiplied by $24. Include any image-input charge only after checking its current CometAPI quote. Add the components across all intended generation and editing requests. Until you have a defensible usage estimate for that workload, leave the per-image total open rather than inserting a made-up token count.

This also makes the comparison fairer. A resolution-tier quote can establish the listed charge for a generated Nano Banana 2 image at that tier; it cannot prove which model costs less per accepted asset. GPT-Image-2.5’s $24/M output rate cannot be compared directly with a $0.0536 image price, nor can a lower-looking rate prove savings. Quality choices, edits, and the number of attempts affect the production budget.

For a fixed per-asset planning line, start with Nano Banana 2’s stated resolution tier. For a GPT-Image-2.5 plan, record the intended quality and obtain a representative token-usage estimate before converting its rates into dollars per image. These figures come from the dated CometAPI article, not a live checkout quote, so recheck the current model prices before committing a budget. The article’s separate discussion of Google’s Batch pricing does not establish a discounted CometAPI Batch price.