How Fast and How Cheap Is AI Image Generation? 2026 Data From 6,369 Images

Short answer: under a minute and under ten cents. The Oakgen AI Image Index 2026, based on 6,369 image generations from December 15, 2025 to November 10, 2026, found that 96.3% of jobs completed. The median image cost 14 credits, about $0.08, and took 42 seconds.

AI images are now cheap and dependable enough that the real cost is your time choosing the right one. Picking the model matters more than any other setting. The slowest common model took more than four times as long as the fastest.

The full dataset, with every sample size, is on the AI image statistics hub.

How often does AI image generation succeed?

Image jobs, December 15, 2025 to November 10, 2026 (n = 6,369).

• 96.3% of image jobs completed.
• 3.4% of image jobs failed.
• 1.0% of image jobs were blocked by content policy.

That is far more reliable than video. On the same platform, 11.5% of video jobs failed (see how often AI video generations fail). Images are smaller, faster jobs with fewer steps to go wrong.

How much does one AI image cost?

These figures use completed image jobs with a credit charge (n = 5,391).

• The median image cost 14 credits.
• The mean image cost 27.1 credits.
• At the add-on rate (180 credits per dollar), the median image is about $0.08.
• On the Basic plan ($19 for 4,000 credits), the median image is about $0.07.
• On the Creator plan ($99 for 30,000 credits), the median image is about $0.05.

Plan rates are as recorded in the dataset on November 10, 2026. The pricing page has current plans.

How long does AI image generation take?

• The median completed image took 41.7 seconds (n = 6,133).

Speed varies by model much more than cost does.

Each column has its own sample. Times use completed jobs, cost uses completed and charged jobs, and completion and share use all jobs. "Not in cost cut" means the model was outside the top models in the cost breakdown, not that it was free.

Which AI image models do people use most?

Share of all image jobs (n = 6,369).

• Nano Banana Pro edit ran 16.1% of image jobs.
• Nano Banana 2.1 edit ran 15.1% of image jobs.
• Nano Banana Pro text ran 10.7% of image jobs.
• FLUX 2 Pro text ran 6.9% of image jobs.
• FLUX 2 Max text ran 6.6% of image jobs.
• GPT Image 2.5 Sunburst edit ran 6.2% of image jobs.

Edits were as common as text prompts. 49.9% of completed, charged jobs were edits that started from an image (n = 5,391). People bring a photo and change it more often than they start from nothing.

Which AI image model should you pick?

Here is how we would use the table.

1. For speed, use Nano Banana 2.1. About 31 seconds, 14 credits and close to 100% completion.
2. For the lowest cost, use FLUX 2 Pro. 6 credits, about $0.03, and every job finished.
3. For edits that must look polished, try Nano Banana Pro. It costs about twice as much, at 30 credits.
4. Expect GPT Image 2 to be slow. Its edit model took a median of 142.6 seconds and finished 76.8% of jobs (n = 272). Save it for jobs where its look is worth the wait.

For a job-by-job guide, read the best AI image generators by use case.

How does this fit the wider market?

Image generation runs at huge scale. OpenAI said more than 130 million users created more than 700 million images in the first week of GPT-4o image generation (OpenAI, April 23, 2025). Canva said its AI products had been used more than 27 billion times, with usage tripling over the past year (Canva, April 13, 2026).

Neither publishes speed, cost per image or failure rates. That is the gap a per-job dataset can fill, even a small one.

How we measured this

• Dataset. The Oakgen AI Image Index 2026 covers every image job on Oakgen from December 15, 2025 to November 10, 2026 (n = 6,369).
• Time. Generation time is from job creation to its completion stamp. Where that stamp is missing, we use the last update, which can only make the time longer. Treat those cases as an upper bound.
• Cost. Cost uses completed jobs with a credit charge above zero. Missing charges were filled from the credit ledger. Medians use the database's approximate median.
• Edits. A job counts as an edit when its model is an image-to-image model.
• Privacy. Aggregate counts only. No prompts, images or user-level data.
• Suppression. Any cut with fewer than 100 jobs is not published, including the share of each failure type.

What are the limits of this data?

• Single platform. Oakgen's prices and model menu shape these numbers.
• Uneven months. Most jobs came from August to November 2026. Early months had fewer than 100 jobs each.
• Queue time counts. Time includes waiting in a provider's queue, which changes with demand.
• Small samples per model. Some models sit just above 100 jobs.

See all cuts on the AI image statistics hub.

Frequently asked questions

How often does AI image generation fail?

Rarely. 96.3% of 6,369 image jobs completed and 3.4% failed. Content policy blocks ended 1.0% of all image jobs.

What is the cheapest AI image model on Oakgen?

Among models with at least 100 charged images, FLUX 2 Pro text-to-image had the lowest median at 6 credits (about $0.03), followed by GPT Image 2.5 Flare at 7 credits.

Are most AI images made from text or from another image?

About half and half. 49.9% of completed, charged image jobs were edits that started from an image (n = 5,391).

Sources

1. Oakgen AI Image Statistics 2026
2. OpenAI: Introducing our latest image generation model in the API (2025-04-23)
3. Canva: The next era of Canva (2026-04-13)