GPT Image 2 Pricing: 2.6 vs 2.5 Tested on 46 Images

OpenAI has added GPT Image 2.6 to the API as two models, gpt-image-2.6-flare and gpt-image-2.6-sunburst, at exactly the same per-token rates as GPT Image 2.5 and GPT Image 2. On the pricing page, nothing changed. On the invoice, something did: GPT Image pricing is set per token, and the new models spend a different number of tokens on the same picture.

So on 7 November 2026 we generated 46 images from our own WordPress server with our own OpenAI key: the same blog featured image on 2.6 Flare, 2.6 Sunburst, 2.5 Flare, 2.5 Sunburst, GPT Image 2 and GPT Image 1 Mini, at every quality setting, twice each. We also tried an edit, a GIF and the two new quality tiers. The whole test cost $1.35.

The short answer: at high quality, a 1536×1024 image costs $0.041 on 2.6 Flare and $0.042 on 2.5 Flare, so it’s a wash. At low and medium quality, 2.6 costs three to four times as much: $0.020 against $0.005 at low. And quality: "auto", which is what most plugins send, picked low on every 2.x model we tried. If your site generates featured images on autopilot, the model ID is now the biggest number on the bill.

GPT Image 2 pricing: the rate card

These are the rates on OpenAI’s API pricing page, read on 7 November 2026, in US dollars per million tokens.

Three things on that page matter more than the numbers themselves:

• 2.6 has no Batch price. Every other current image model gets 50% off through the Batch API. The 2.6 model pages list only the Images API (v1/images/generations and v1/images/edits); Batch and the Responses API are marked “Not supported”.
• Cached input rates don’t help 2.6 either. OpenAI says cached image rates apply only to images made through the Responses API, which 2.6 doesn’t support.
• OpenAI publishes no price per image. It points you to a calculator. What you pay depends on how many output tokens a model decides an image needs, and that’s what we measured.

OpenAI describes the two 2.6 models as Flare “for fast iteration” and Sunburst “for detailed visual work”. Both accept a new complexity field, which OpenAI says carries no separate fee.

How we tested it

All requests came from a PHP script on our WordPress server, calling the Images API directly with the OpenAI key our site uses. We copied the request our own plugin’s content generator sends for a blog featured image: 1536×1024, PNG, one image, and whatever quality the site owner picked, defaulting to auto. The prompt asked for an editorial illustration for a post called “How to Cut WooCommerce Shipping Costs”, with a parcel on a kitchen scale, a laptop showing a checkout page and the headline “Cut Shipping Costs” across the top. Text in an image is where cheap image models usually fail, so it’s a useful test.

• Quality grid: low, medium, high and auto on 2.6 Flare, 2.6 Sunburst, 2.5 Flare and 1 Mini; auto on 2.5 Sunburst and GPT Image 2. Every setup ran twice, which made 36 images.
• Probes, once each: 2.6 Flare at the new xhigh and max tiers, at complexity low and medium, and at 1024×1024; 1 Mini at 1024×1024; and a GIF.
• Edits: one edit of a 2.6 image, run on 2.6 Flare at low and high and on 2.5 Flare at high.

That came to 48 requests, of which 46 returned an image. The Images API returns a token count with every response, so the costs below are the reported tokens multiplied by the published rates, not an invoice reconciliation. Token counts were identical on both runs of every setup, so the per-image prices are fixed for a given model, size and quality. The timings are one morning’s sample from one server in the US.

What one image cost

The prompt itself was 73 text tokens, which is $0.0004 at $5 per million. Every price above is almost entirely image output tokens at $30 per million ($8 per million for 1 Mini).

The 2.6 surprise is at the bottom of the range. 2.5 Flare draws a low-quality 1536×1024 image in 158 tokens; 2.6 Flare uses 664 for the same request, so it costs four times as much on an identical rate card. At medium the gap is 1,132 tokens against 343. At high, the two land within 19 tokens of each other. OpenAI hasn’t said why; the practical effect is that “low” on 2.6 costs about what “medium” or better costs on 2.5.

The auto surprise is bigger. When you send quality: "auto", the response reports which tier the model chose. Every 2.x model chose low for our featured image, on both runs. GPT Image 1 Mini chose high, which made it the most expensive way to draw this picture in the whole test at $0.0498, more than 2.6 Flare at high. If you’ve been running 1 Mini as the “cheap” model with the default setting, you’ve been paying for its top tier.

Two settings made no difference to cost. The complexity field at low and medium returned the same 664 tokens as the default on 2.6 Flare, which matches OpenAI’s “no separate fee” line. And Flare against Sunburst on 2.6 cost the same at low and medium, and 19 tokens apart at high.

One sizing oddity is worth knowing: a square 1024×1024 image on 2.6 Flare at auto cost $0.0251 (824 tokens), more than the larger 1536×1024 at $0.0203 (664 tokens). On 1 Mini it went the other way: 1024×1024 at auto cost $0.0334 against $0.0498 for 1536×1024. We ran the square size once per model, so take that as a pointer to test your own sizes, not a rule.

What you get for the money

Price only matters if the picture is usable, so here are the first runs side by side.

Every 2.x image spelled the headline correctly, at every quality. Low quality on both 2.5 and 2.6 looks like a finished photograph at blog size. The differences are in detail: 2.6 rendered the laptop’s checkout screen in larger, more readable type at low quality, and added more small props as quality rose. On 2.5, the screen text gets smaller and fussier, but you’d only notice at full size. For a featured image shown at 800 pixels wide, 2.5 Flare at low is good enough, and it’s a quarter of the price.

GPT Image 1 Mini read “editorial illustration” as a flat cartoon at every tier. Its headline was right, but at low quality the laptop’s button said “CARCHOUT”. GPT Image 2 at auto produced a photographic scene close to 2.5’s. Sunburst’s images were the most polished at high, with a fuller checkout page and cleaner shadows. The new max tier on 2.6 Flare is sharper again, but at $0.165 it costs eight times 2.6 low. That’s hard to justify for anything shown at blog width.

How long each one took

This is where 2.6 Flare earns its name. It returned a high-quality image in about 7 seconds, less than half of 2.5 Flare’s 18.8 seconds at high, and its speed barely changed with quality. xhigh took 16.0 seconds and max 22.6. Sunburst went the other way, at 23 to 43 seconds, so it’s the slowest current option. GPT Image 2 at auto took 13.2 seconds and 2.5 Sunburst at auto 11.8.

Speed matters most when a person is waiting, for example a chatbot that draws on request or an editor clicking “regenerate”. For a featured image generated in the background while a post saves, 10 extra seconds rarely matters, and the 4x price difference at low does.

Editing and GIFs

Edits go through v1/images/edits with the source image uploaded. We took the 2.6 Flare high-quality image and asked for a red parcel and a new headline, “Ship for Less”.

All three edits kept the scene intact: the same scale reading, the same checkout screen, the same mug and plants, with only the parcel and headline changed. The uploaded image counted as 1,536 input tokens at $8 per million, so an edit costs about $0.012 more than generating the same image fresh. The low-quality edit was indistinguishable from the high one at blog size, for $0.021 less.

GIF output is new in this family. Our first try, 1024×1024, failed straight away with a 400: “No supported sheet fits 9 GIF frames at size ‘1024×1024′”. The frames are drawn as one large sheet, and nine 1024-pixel frames make a sheet bigger than the model will draw. A frames parameter returned “Unknown parameter” on the Images API. At 512×512 it worked: 9 frames, 200 ms each, 724 KB, $0.0356, 9.8 seconds.

The model chose its own animation. It counted the shipping line down from $12.00 to FREE across nine frames, and the totals added up on every frame. That’s a fun touch, but at 724 KB it’s too heavy for a featured image. It works better as a one-off for social posts.

What it costs per 1,000 images

This is the measured per-image cost multiplied by 1,000. It’s arithmetic, not a separate test.

A site that publishes one post a day uses about 365 featured images a year, which is $1.86 on 2.5 Flare at auto, $7.40 on 2.6 Flare at auto, and $18.18 on 1 Mini at auto. Those are small numbers, which is the real point: for a blog, picture quality and speed matter more than price. The price gap only bites at volume, for example product image variants across a WooCommerce catalogue, or a chatbot that draws for visitors. Bulk back-catalogue jobs are where the missing Batch discount on 2.6 hurts most. On list prices, 2.5 Flare at high through Batch costs about half of 2.6 Flare at high, for a picture of similar quality.

GPT Image 1 Mini shuts down on 1 December

OpenAI’s deprecations page lists 1 December 2026 as the shutdown date for gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest, with gpt-image-2 as the named replacement. Base gpt-image-1 already shut down on 23 October. Code that still calls those IDs after 1 December will get errors, not images. We covered how to plan for this kind of change in what OpenAI model deprecations mean for a WordPress chatbot.

Our test changes the usual advice slightly. If you’re on 1 Mini at the default setting, moving to GPT Image 2 or 2.5 Flare at auto cuts the cost per image by about 90%, from $0.0498 to $0.0051, because auto stops picking the top tier. You’ll also get a photographic image rather than a cartoon, so check that the new style suits your site before you switch a whole catalogue. If you set 1 Mini to low on purpose to save money, 2.5 Flare at low costs about 50% more ($0.0051 against $0.0033). Nothing current is cheaper per image than that.

Which GPT Image model to use on a WordPress site

One setting matters more than the model: don’t assume auto means “sensible”. It meant the cheapest tier on every 2.x model and the most expensive tier on 1 Mini. If cost or quality matters, set the quality explicitly.

What this means for MxChat sites

MxChat’s AI content generator makes featured images with the same request we tested: 1536×1024, PNG, at the quality you choose, which defaults to auto. Its OpenAI model list offers GPT Image 2.5 Flare, 2.5 Sunburst and GPT Image 2. 2.6 isn’t in that list yet, so this test is what you’d be choosing between if you call the API yourself. If a saved setting still names gpt-image-1-mini, gpt-image-1.5 or chatgpt-image-latest, MxChat sends 2.5 Flare instead once the 1 December shutdown passes, so images keep coming. On the default setting, that’s the $0.0051 row above.

Images are a small line next to what the chatbot itself costs to run. For that side, see what the OpenAI Batch API saves on chatbot costs and how OpenAI prompt caching is billed. Our Gemini Omni pricing test does the same exercise for product video. MxChat runs on your own API keys, so whatever you pay OpenAI is the whole cost; MxChat Pro and the documentation cover the setup.

Frequently asked questions

How much does GPT Image 2.6 cost per image?

For a 1536×1024 image we measured $0.0203 at low, $0.0343 at medium, $0.0410 at high (Flare) or $0.0415 (Sunburst), $0.0741 at xhigh and $0.1650 at max. The rates are $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens.

Is GPT Image 2.6 more expensive than 2.5?

The rate card is identical, but 2.6 uses more output tokens at low and medium quality. In our test, low cost $0.0203 on 2.6 Flare against $0.0051 on 2.5 Flare, and medium cost $0.0343 against $0.0107. At high, the two cost the same.

What is the difference between Flare and Sunburst?

OpenAI positions Flare for fast iteration and Sunburst for detailed work. They cost the same in our test. Flare took about 7 seconds at high and Sunburst 43 seconds.

What quality does “auto” pick?

For our 1536×1024 featured image, auto picked low on 2.6 Flare, 2.6 Sunburst, 2.5 Flare, 2.5 Sunburst and GPT Image 2, and high on GPT Image 1 Mini. The response tells you which tier was used.

Does GPT Image 2.6 work with the Batch API?

No. OpenAI’s 2.6 model pages list Batch as not supported, and the pricing page shows no Batch price for 2.6. 2.5 and GPT Image 2 get 50% off through Batch.

How much does a GPT Image edit cost?

Our 1536×1024 edit cost $0.0324 at low and $0.0531 at high on 2.6 Flare. The uploaded image added 1,536 input tokens, about $0.012 on top of a fresh image.

Can GPT Image 2.6 make GIFs?

Yes, with output_format: "gif". A 512×512 GIF of 9 frames cost $0.0356 and weighed 724 KB. At 1024×1024 the request was rejected because nine frames didn’t fit.