GPT Image 2.6 API pricing: image costs and GIF/SVG billing

GPT Image 2.6 Flare and Sunburst have the same published OpenAI token rates: $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens. That does not make every image, GIF or SVG the same price. The amount of work, input material and output consumption still matters. SVG adds another billing component: the model that authors the vector document.

Checked November 5, 2026. This is a source-based pricing and access guide, with worked arithmetic rather than a claimed generation benchmark. We have not made paid Image 2.6 generation requests for this article. Ofox publishes this guide; official OpenAI prices below are not an Ofox quote.

Which Image 2.6 can you use?

OpenAI announced Images 2.6 on November 4, with ChatGPT, Codex and API access. The two API identifiers are gpt-image-2.6-flare and gpt-image-2.6-sunburst. Use the identifier required by the provider you actually call, not a display name copied from a blog title.

The Ofox check covered all 126 records returned by its public catalog and the authenticated model list available to our account. It establishes what those sources listed at that time, not a permanent statement about the service. The currently listed GPT Image 2.5 Flare and Sunburst are separate versions. Their availability does not prove support for 2.6 GIF or SVG requests. Follow Ofox model documentation for its own access path.

Official token rates, with the units intact

The Flare model page and Sunburst model page list these rates in US dollars:

A cached rate being listed is not proof that your request received it. The current image guide’s paragraph about cached-input endpoint eligibility explicitly names earlier Image 2 and 2.5 models. We do not extend that statement to 2.6. Use confirmed billing data for the route and request you ran; leave unknown cache savings out of a forecast.

Nor does the table say that Flare and Sunburst consume equal tokens for the same task. Model choice, dimensions, quality and the number of attempts all affect a useful budget. Check the current OpenAI pricing page before committing to a batch.

Calculate one ordinary image request

For a request with no confirmed cache discount and no additional authoring or conversational model charge, use:

USD = (text_input_tokens × 5
     + image_input_tokens × 8
     + image_output_tokens × 30) / 1,000,000

Hypothetical example, not a measured Image 2.6 response: 120 text input tokens, no image input and 1,000 image output tokens cost (120 × 5 + 1,000 × 30) / 1,000,000 = $0.0306 under these rates. Adding 1,024 image input tokens would add $0.008192, bringing the example to $0.038792. Those token counts illustrate the calculation; they do not predict how many tokens your image will consume.

Keep text and image inputs separate. If a response provides an aggregate input total and its breakdown, do not bill both the total and the breakdown. If the detail is missing, record the gap rather than silently treating all input as text. Where cached tokens are confirmed, separate them from uncached tokens; do not count them twice.

The official output calculator helps estimate the image-output component using explicit model, quality and size. It is not the entire bill for a request with reference images or other model work. An auto setting is also less useful for a repeatable estimate than an explicit setting.

Budget for accepted results, not just API responses

Suppose a small pilot gives you an average billed attempt of $0.0306 and an acceptance rate of 80%. These remain hypothetical assumptions until you measure them. To obtain 100 accepted images, the expected number of attempts is 100 / 0.8 = 125, giving an illustrative generation budget of $3.825. At 50% acceptance it becomes 200 attempts, or $6.12, even though the token price is unchanged.

Acceptance means meeting the task: the right lettering, an intact product label, correct dimensions and no unacceptable layout changes. An HTTP 200 response only means the request returned successfully. Your own discarded but successfully generated outputs may still contribute to your bill. This is different from making a general claim about how a provider bills failed HTTP requests.

For a first pilot, record model, endpoint, request settings, input count, elapsed time, returned usage, actual charge and acceptance reason together. Keep rejected samples as well as winners. Do not use a single quick request as proof of stable latency or a fixed production price.

GIF and SVG need separate estimates

The image-generation cost guide describes the SVG authoring charge but does not name a universal authoring-model rate in that section. Do not quote an all-in SVG price by multiplying everything by $30 per million. Likewise, multiplying a still-image estimate by a GIF’s frame count is not a verified billing formula. Use the provider’s actual usage and billing records for those jobs.

Our GIF workflow explains how to save and inspect the animation. The SVG workflow checks whether the result contains usable vector elements and how to review it before handoff. Both include runnable local helpers and explicitly distinguish documentation-based requests from live API tests.

Does lower complexity make it cheaper?

quality controls rendering detail; complexity controls planning. OpenAI documents low, medium and high complexity, defaulting to high when omitted, and says there is no separate fee for adjusting this setting. That does not establish that every setting produces the same bill or that low always meets the task.

For a straightforward change, compare low complexity with your baseline while keeping the input, size and quality unchanged. For dense text or several spatial constraints, inspect whether less planning causes more rejected outputs. A lower per-attempt cost is not a saving if you need enough additional attempts to erase it. This article has not measured that trade-off for 2.6.

Before funding or migrating a workflow

1. Confirm that the exact model is accessible on your chosen provider. Keep OpenAI and Ofox credentials and base URLs separate.
2. Start with one request and a small explicit spending limit. Never expose an API key in a webpage or shared screenshot.
3. Save usage and compare it with the account’s billed amount. Keep estimates distinct from settled charges.
4. Validate the actual deliverable. For GIF, check frames and motion; for SVG, inspect structure and editability.
5. Expand only after the sample meets your acceptance rule. Treat a 2.5-to-2.6 switch as a new regression check rather than a name replacement.

For an existing Ofox 2.5 workflow, the 2.5 pricing guide contains historical measurements specifically labeled for that version. They are useful for understanding the recording method, not for inventing 2.6 results. Consult Ofox rate-limit documentation separately from pricing: funding an account does not itself prove a particular concurrency allowance.