Budgeting Sunburst on CometAPI: token rates are not a fixed per-image price

If you are turning CometAPI’s Sunburst listing into an image-credit allowance, I would stop short of pricing each image at $0.028. The page calls that an estimated first-request cost, not a fixed per-image tariff. It is enough to start a budget conversation, but not enough to promise how many finished images a user’s credits will buy.

This reading is based on CometAPI’s GPT Image 2.5(sunburst) listing, captured September 13, 2026. All prices and model descriptions below are attributed to that listing; they are not independently verified provider pricing or results from a billing test.

Three numbers that answer different questions

What CometAPI displays What it supports for planning
$4 per million input tokens and $24 per million output tokens A preliminary token-rate comparison, subject to clarifying the input category
Comparison rates of $5/M input and $30/M output, with a 20% discount The displayed reductions are both 20%; this does not establish savings for every workflow
$0.028 for the first request, $0.28 for ten, $0.42 for fifteen A request-cost estimate, not a disclosed image-price schedule

The ten- and fifteen-request figures are simply linear multiples of the first. They add no information about token consumption or how many images a request delivers. Multiplying $0.028 by a monthly image allowance would carry those unknowns into the product’s credit system rather than resolve them.

The input category needs clarification

The same page’s specifications distinguish prices it labels “official”: $5/M for text input, $8/M for image input, and $30/M for image output. Its CometAPI pricing comparison collapses input into one category at $4/M.

That leaves a practical question for an editing product: what CometAPI rate applies to reference-image input? I would neither assume the $4/M input rate covers images nor calculate an unlisted discounted image-input rate from the 20% headline. The listing’s comparison alone does not settle that mapping.

This matters to the scope of the budget, not just its arithmetic. CometAPI describes Sunburst as supporting generation and editing, with quality settings and configurable dimensions. Those descriptions give a team reasons to define separate request scenarios; they do not establish the cost of those scenarios.

Before converting the estimate into credits

The first-request estimate is not tied to disclosed input and output token counts, text-only versus reference-image input, reference-image quantity and dimensions, output dimensions, quality setting, or output image count. A sample request elsewhere on the page is not an explicit explanation of that estimate.

My recommendation is to ask CometAPI for the estimate’s assumptions and the modality-specific billing mapping, then build a small budgeting worksheet:

  • Generation scenario: record the prompt, intended output dimensions, quality setting and image count.
  • Editing scenario: additionally record reference-image quantity and dimensions, plus the planned number of editing rounds.
  • Usage reconciliation: for representative requests, record reported usage, resulting image count and observed charges, where available. Keep unresolved billing categories visible rather than filling them with guessed rates.
  • Workflow allowance: ask how editing rounds, failures and retries are billed before deciding what a user-facing credit covers. Do not assume either free retries or a charge for every failed attempt.

For the interface, I would keep an estimated request cost distinct from a finished-image allowance until those checks are complete. If the product offers a fixed credit package, its rules should explain whether a credit buys an initial generation, an editing round, or a defined bundle—not silently inherit the listing’s first-request estimate.

Bottom line: CometAPI’s Sunburst page supports an initial rate comparison. Setting fixed per-image credits needs the missing workload assumptions and representative usage evidence first. The question to resolve is not whether $0.028 scales arithmetically; it is what that estimate actually buys.

For a hosted-versus-local cost comparison, I would add an acceptance denominator to this worksheet. The Sunburst listing’s $0.028 estimated request is not a useful break-even target for local inference: it is neither a measured workflow cost nor a count of outputs that meet the product’s requirements.

My proposed comparison would use total trial cost divided by accepted final assets, with the acceptance rules fixed before either route produces results. For an editing trial, those rules could include completing the requested change while preserving specified reference details. Record rejected outputs and abandoned workflows as well as successful ones; otherwise a route can appear cheap simply because the comparison excludes its unsuccessful work. If no assets pass, report that outcome rather than a per-asset cost.

On the Sunburst side, use reconciled charges once the input-billing ambiguity is resolved—not the headline estimate multiplied by attempts. For a local candidate, separately record the exact model revision, quantization, runtime, hardware, GPU time and an explicit allocation for idle capacity. Show setup and review labor separately so readers can see whether they are comparing inference expense or the wider workflow budget.

This is a proposed evaluation design, not evidence that a local model matches Sunburst or costs less. A cheaper generated image and a cheaper accepted editing result are different claims; I would require the same acceptance test before using either to justify a hosting decision.