GuidesOneMux

GPT Image 2.6 Sunburst API Costs: A Reproducible Evaluation Budget

Estimate Sunburst image-generation costs with explicit token assumptions, acceptance criteria, spending limits, and a repeatable test plan.

AI

Summary · TL;DR

A practical cost guide for evaluating GPT Image 2.6 Sunburst, including source-reported token rates, an illustrative budget, and a repeatable acceptance test.

Key Takeaways

  • The supplied source reports direct OpenAI rates of $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens; verify current rates before a paid test.
  • OneMux's published snapshot did not include usable input or output prices for Sunburst, so it cannot support a numerical OneMux workload estimate.
  • Budget for rejected outputs and repeat attempts, not just successful API responses.
  • Keep test settings, acceptance rules, returned usage, and billed charges together so another evaluator can reproduce the calculation.
On this page (8)

Start with a capped pilot, not a per-image quote

For a repeatable GPT Image 2.6 Sunburst evaluation, set a spending limit, fix the request settings, and judge every returned image against the same acceptance rules. Estimate the budget from input and output token assumptions, then replace those assumptions with observed usage and charges. A successful API response is not necessarily an acceptable image.

This cost guide is published by OneMux, which offers model access through an OpenAI-compatible API. The numerical example below uses source-reported direct OpenAI rates, not OneMux prices. It is illustrative: no paid Sunburst requests were made for this article.

Prices and comparable units

The supplied Ofox pricing guide reports the following OpenAI rates for Sunburst, checked November 5, 2026. They are US dollars per one million tokens, not prices per image. Verify current provider pricing and access before funding a test.

Direct OpenAI billing categorySource-reported priceUnit
Text input$5.001 million tokens
Image input$8.001 million tokens
Image output$30.001 million tokens
Cached text input$1.251 million tokens
Cached image input$2.001 million tokens

A listed cache rate does not establish that a particular Sunburst request qualifies for it. Forecast without cache savings unless the route's documentation and billing data confirm them. If cached usage is confirmed, separate it from uncached usage rather than counting the same tokens twice. Check applicable tiers and conditions with the provider; the example below assumes no tier adjustment.

The OneMux catalogue snapshot captured November 15, 2026 provides a separate, incomplete comparison:

ModelOfficial inputOfficial outputOneMux inputOneMux outputPublished price ratio
GPT Image 2.6 SunburstUnknownUnknownUnknownUnknown0.11699999999999999

The catalogue does not supply token-category prices or units for those unknown fields. Its published ratio is reproduced exactly, but without usable underlying prices and conditions it is not enough to calculate a OneMux charge or claim a particular saving. Unknown is not zero. Check current OneMux prices and the model's terms before comparing routes.

Calculate a workload using explicit assumptions

Assume a direct OpenAI request uses 120 text input tokens, no image input, and 1,000 image output tokens. Assume no confirmed cache discount or additional model charge. These are invented test inputs for arithmetic, not measured Sunburst consumption or a prediction for a given image size.

ComponentCalculationIllustrative cost
Text input120 × $5 / 1,000,000$0.0006
Image output1,000 × $30 / 1,000,000$0.0300
Total per attempt$0.0006 + $0.0300$0.0306

If the same hypothetical request also used 1,024 image input tokens, that component would add 1,024 × $8 / 1,000,000 = $0.008192, for a total of $0.038792. Reference images therefore belong in the estimate, even when the prompt is short.

Now suppose the average attempt costs $0.0306 and 80% of outputs meet the acceptance rule. Obtaining 100 accepted images would take an expected 100 / 0.80 = 125 attempts, costing 125 × $0.0306 = $3.825. At 50% acceptance, the expectation is 200 attempts and $6.12. Neither figure is a spending ceiling: actual attempts and token usage can vary.

Run a bounded, repeatable evaluation

Before testing, confirm access to the exact model and endpoint, current rates, credentials, and the supported image request parameters. OneMux compatibility for a particular workflow depends on its documented endpoint and parameter support; this guide does not claim an integration test. Keep API keys out of shared logs and screenshots.

Use this illustrative protocol for a raster-image evaluation:

  1. Set hard limits. Cap the pilot at 20 attempts and set a provider-side spending limit if available. Stop earlier if the limit is reached, usage cannot be reconciled, or the estimated remaining budget is insufficient. Do not rely on an expected cost as a hard cap.
  2. Freeze the request. Record the exact model identifier, provider and endpoint, prompt, reference-image count, dimensions, quality, and any other supported settings. Avoid automatic settings where explicit ones are available and relevant.
  3. Prepare cases before generating. For example, use five prompts covering legible short text, object placement, preservation of a reference label, a simple edit, and a visually dense layout. Run four attempts per case in a predetermined order; do not selectively rerun only weak cases within the initial 20-attempt comparison.
  4. Write pass criteria in advance. For the short-text case, specify the exact wording and require every character to be legible and correct. For placement, specify the required relative positions. For a reference-label edit, state which label elements must remain unchanged. Apply the same rule to all attempts in each case.
  5. Keep every result. Record pass or fail and the reason, including for discarded images. An HTTP success is not a quality pass; do not infer a provider's treatment of failed HTTP requests from that distinction.

For each attempt, retain the case ID, settings, timestamp, response status, input and output usage breakdown if returned, elapsed time, acceptance decision, and charge when available. Store the output or a secure reference to it so another reviewer can check the decision. Reconcile the pilot total against the provider's billing record; an estimate is not a settled charge.

If usage reports only an aggregate input total, do not add it again to a detailed breakdown or assume all of it is text. Mark the component costs unknown until the provider supplies enough detail. If the model or a required parameter is unavailable, stop and document the limitation rather than silently substituting a different model or setting.

Interpret the results without overclaiming

Divide the observed pilot charge by the number of accepted images to obtain an observed cost per accepted image. Report the sample size, acceptance criteria, settings, and any unassigned charges alongside it. A 20-attempt pilot is useful for finding obvious failures and estimating a follow-up budget; it does not establish stable production quality, latency, or cost.

To compare a different setting or model, rerun the same cases and acceptance rules, changing only the intended variable. The source reports identical published token rates for Flare and Sunburst, but that does not establish equal token consumption or equal output quality. Neither comparative performance nor savings have been measured here.

GIF and SVG work needs a separate plan. The source says animation workload depends on factors including frames and dimensions, while SVG can include a separate authoring-model charge. Do not multiply the still-image estimate by frame count or apply the image-output rate to the entire SVG job as though either were a verified billing formula. Record actual usage and charges for each component instead.

Conclusion

Use the published token rates to build an explicit starting estimate, then use a capped pilot to measure the cost of outputs your task actually accepts. Keep direct-provider and OneMux prices separate until comparable, current billing units and conditions are available.

Sources

  • Ofox: GPT Image 2.6 API pricing — source-reported OpenAI rates, access notes, and GIF/SVG billing cautions; the guide says it did not make paid Image 2.6 generation requests.
  • OneMux pricing — current published pricing; the OneMux catalogue snapshot cited above was captured November 15, 2026.

Published price snapshot

Captured 2026-11-15T11:00:05.734Z from the OneMux published catalogue. Prices retain website billing units; this is not a live benchmark. Tiered, cached or time-based billing may differ. Unknown values are not zero.

ModelOfficial inputOfficial outputOneMux inputOneMux outputPublished price ratio
GPT Image 2.6 SunburstUnknownUnknownUnknownUnknown0.11699999999999999

Current OneMux prices. Compare workload costs only when input/output and billing units match; see each model page for tier conditions.

FAQ

How much does one GPT Image 2.6 Sunburst image cost?

There is no fixed per-image cost established by the supplied source. As an illustration at its reported direct OpenAI rates, 120 text input tokens and 1,000 image output tokens would cost $0.0306, excluding image inputs, cache adjustments, and other model charges. Actual consumption and provider pricing may differ.

Can I calculate the same workload's OneMux price from the published snapshot?

No. The snapshot lists Sunburst input and output prices as unknown. Its published price ratio alone does not provide enough information about underlying prices, units, or conditions to calculate a workload charge.

Why measure cost per accepted image?

A returned image may fail the task's acceptance criteria. Dividing the observed charge by accepted outputs includes the effect of rejected attempts, making the result more useful for planning the same workflow.

Are Sunburst and Flare equally expensive to use?

The supplied source reports identical published OpenAI token rates for the two models. It does not establish that they consume the same number of tokens or achieve the same acceptance rate for a given task.

Related articles