GPT Image 2.6 API Pricing: Every Quality and Resolution Tier

The Full Price Table at a Glance

Every price below is USD per image, read from the live /api/pricing endpoint. Flare and Sunburst share the same numbers. Find your row, then pick your column.

• low: 1K 0.023 · 2K 0.042 · 4K 0.075

• medium: 1K 0.041 · 2K 0.075 · 4K 0.133

• high: 1K 0.049 · 2K 0.090 · 4K 0.160

• xhigh: 1K 0.087 · 2K 0.160 · 4K 0.283

• max: 1K 0.195 · 2K 0.360 · 4K 0.636

That is 15 quality × resolution tiers, from 0.023 at low 1K to 0.636 at max 4K — nearly a 28× spread inside one model. Two output formats sit outside this grid: SVG adds a flat 0.030 to the same-tier PNG price, and GIF is billed by quality only, ignoring resolution entirely (low 0.070 through max 0.598). Image-to-image reuses the same table and adds 0.012 per reference image. The rest of this post walks through each axis so you can pick a tier on purpose instead of by habit.

What Each Quality Tier Actually Costs You

Read the table by column and the quality ladder looks like this at 1K:

• low → medium: 0.023 → 0.041, a step of +0.018

• medium → high: 0.041 → 0.049, +0.008

• high → xhigh: 0.049 → 0.087, +0.038

• xhigh → max: 0.087 → 0.195, +0.108

The same shape holds at 2K and 4K. Two things stand out. First, medium → high is the cheapest upgrade in the whole ladder — at 1K you pay 0.008 for a full quality level. If a draft is almost there, high is usually the better spend than another low attempt. Second, xhigh → max is the most expensive step by a wide margin: at 1K, +0.108 is more than the whole low → xhigh climb (+0.064). max is a deliberate choice for final output, not a default.

These tiers map to the quality parameter (low / medium / high / xhigh / max). Note that quality is independent of complexity (low / medium / high, default high), which controls the planning level of the generation — changing one does not move the other's price.

How Resolution Multiplies the Bill

Resolution is the resolution parameter: 1K (default), 2K, 4K. Within each quality level, the jumps are not uniform.

• low: 0.023 → 0.042 → 0.075. The 2K step is ~1.8× the 1K price; 4K adds about 79% on top of 2K.

• medium: 0.041 → 0.075 → 0.133. The same ~1.8× at the first step.

• high: 0.049 → 0.090 → 0.160. 2K is ~1.8× of 1K; 4K is ~1.8× of 2K.

• xhigh: 0.087 → 0.160 → 0.283. The same ~1.8× cadence.

• max: 0.195 → 0.360 → 0.636. 4K is ~3.3× the 1K price at this tier.

The steepest absolute jump sits at the top: max 2K → max 4K is +0.276 per image, the price of twelve low 1K images. In relative terms, every 1K → 2K step lands between 1.82× and 1.85×, so moving up one resolution step costs close to double at any quality. Practical rule: compare across the grid before committing. High 4K (0.160) is cheaper than max 1K (0.195), and medium 4K (0.133) is cheaper than max 1K too — so if you mainly need size, a lower quality tier at higher resolution can cost less than the reverse.

SVG: Same Tier Plus a Fixed Surcharge

output_format: "svg" is available on text-to-image only. It is billed as the same-tier PNG price plus a flat 0.030.

• low 1K SVG: 0.023 + 0.030 = 0.053

• max 4K SVG: 0.636 + 0.030 = 0.666

Because the surcharge is fixed, it is proportionally heaviest at the bottom of the grid: at low 1K it more than doubles the per-image cost. At max 4K it is a small slice of the total. If you are generating vector output in a loop, the surcharge is the line item to watch; if you are producing a handful of final assets, it barely registers. SVG is not available on image-to-image — that path outputs png / jpeg / webp only.

GIF: Priced by Quality, Not Resolution

GIF billing ignores resolution entirely. The price depends only on the quality level:

• low: 0.070

• medium: 0.124

• high: 0.150

• xhigh: 0.266

• max: 0.598

Frame size (frame_size), frame count (frame_count, 2–16, default 9), and frame duration (frame_duration_ms, ≥20, default 200, rounded to 10 ms automatically) do not change the price. A 512×512 nine-frame loop and a 768×768 twelve-frame loop at the same quality cost the same. That makes GIF one of the few places where you can raise frame count for smoothness without touching the bill — though frame_size of 768×768 caps at 12 frames.

The billable event is still gated by validity: the whole stitched frame sheet (not each frame) must keep each side ≤3840, total pixels ≤8,294,400, and aspect ratio ≤3:1. Submissions that break those limits, or use invalid parameters, are rejected at submit time and refunded in full — no charge. GIF is text-to-image only, and aspect_ratio should be omitted or set to auto when generating one.

Image-to-Image and the Per-Reference Fee

Image-to-image uses the exact same quality × resolution table as text-to-image, then adds 0.012 per reference image. You can pass 1–16 source images (jpeg/png/webp).

A worked example: a 4K max edit with 16 reference images costs 0.636 + (16 × 0.012) = 0.636 + 0.192 = 0.828 before any output-format considerations. At the other end, a low 1K edit with a single reference is 0.023 + 0.012 = 0.035.

The per-reference fee is flat, so it scales linearly with how many source images you attach — not with their size. Even a single reference is about a third of the bill on a low 1K edit (0.012 of 0.035), but under 2% on a max 4K edit (0.012 of 0.648). It is worth tracking reference count separately in your cost accounting.

Masking does not add a fee: mask_url (an RGBA PNG matching the first source image's dimensions) and markup_url (a transparent annotation PNG) are mutually exclusive inputs, and neither changes the price. Note that with a mask, areas outside the masked region may still shift slightly. Output on this path is png / jpeg / webp only.

Flare and Sunburst: Identical Price, Different Fit

The four model IDs — gpt-image-2.6-flare/text-to-image, gpt-image-2.6-flare/image-to-image, gpt-image-2.6-sunburst/text-to-image, gpt-image-2.6-sunburst/image-to-image — share identical parameters, usage, and pricing. Every number in this post applies to both lines.

So the choice is not a budget decision. It is a workflow decision:

• Flare is the faster line, suited to rapid iteration and batch generation. It is the recommended default.

• Sunburst leans toward detail; at medium quality and above, multi-turn edits hold character, product, and layout consistency better.

A common pattern is to draft on Flare and switch to Sunburst for the passes where consistency across edits matters — the price does not change when you switch. Because the two lines are priced identically, you can move work between them freely without re-running your cost model.

Picking a Tier for Your Workload

Four rules, derived from the numbers above:

Iteration loops

Stay at low 1K (0.023) or medium 1K (0.041) while you are still exploring prompts. At 0.023 per image, a hundred probes cost $2.30 — cheap enough that trying more variants beats agonizing over one. Bump to high (0.049 at 1K) only when a draft is close, since medium → high is the smallest step in the ladder.

Batch generation

Batch work multiplies every line item. A 1,000-image run costs $23 at low 1K, $41 at medium 1K and $42 at low 2K — one step up on either axis nearly doubles the run. Price the exact cell you need before scaling up. Flare's speed also pays off here — for identical cost, it gets you through the batch sooner.

Final renders

Reserve xhigh and max for output you will actually ship. The xhigh → max step is the most expensive in the ladder (+0.108 at 1K), and max 4K at 0.636 is nearly 28× the cheapest tier. Before committing to 4K max, check whether high 4K (0.160) or max 1K (0.195) gets you what you need.

Edit-heavy pipelines

Reference count is your hidden cost driver. At 0.012 per reference, a 16-image edit adds 0.192 — more than eight times the price of a low 1K output (0.023). Keep reference sets tight unless the edit truly needs them. Pair Sunburst with medium-or-above quality when consistency across rounds matters more than per-image price.

One safety net across all of the above: submissions rejected at submit time for invalid parameters or GIF size limits are refunded in full. You can validate the format rules before optimizing anything else. For per-model usage details, see the Flare text-to-image and Sunburst image-to-image pages, and the full table on the pricing page.