It takes about 4 AI images to keep one

Section 01

1 download for every 4 generations

One image is almost never the end of it. A working session, consecutive generations by the same person in the same tool with less than 10 minutes between them, has a median of 4 generations. 18.6% of sessions stop after 1; 57% stop at 4 or fewer; 1 in 10 runs past 20, and those sessions produce half of all the generations this report counts.

Downloads tell the same story from the other end. In the image generator, people downloaded 1 image for every 4.0 they generated. In Talk to image, 1 for every 4.8; in Spaces, where a workflow generates several images per step, 1 for every 5.7. Across every image tool, 1 for every 4.6. Weight each user the same instead of each image and the typical user sits at 6 generations per download, because the heaviest users generate the most and keep the smallest share. 3 in 4 days a person uses the image generator end with at least one download; 1 in 4 end with nothing kept.

Section 02

The image you keep costs about 300 credits

One generation costs 75 credits in 48% of cases and 50 credits in another 33%. The median is 75; the mean over priced generations, 72. 6% of generations cost 25 credits or less, 9% cost 100 or more.

The list price is the model: Seedream 4.5, Nano Banana, and Seedream 4 at 50 credits; Nano Banana 2, Nano Banana Pro, and Seedream 5 at 75; GPT Image 2 from 15 to 260 depending on quality, 45 at the median and 173 on average. Count every generation, the free ones included, and divide by the downloads: the kept image comes out at 284 credits in the image generator, 331 in Talk to image, 445 in Spaces, 321 across all image tools. The median session costs 250 credits, 3 to 4 generations at 75; 300 if only sessions with a priced generation count. 2 in 3 generations run on unlimited plans, where the credit is a list price, not a charge; on the plans that pay per generation, the mean was 69 credits in February and 99 in August.

Section 03

Each try costs 46% more—the kept image, 9 to 19%

In February 2026 the average generation cost 60 credits at list price. In August, 88, up 46%. Almost none of that is price changes.

At each model’s window price the average would have risen just as much, because people moved to pricier models: Seedream 4.5, at 50 credits, went from 37% of priced generations to 4.5%; Seedream 5 Pro 2K, at 75 credits, went from 0 to 26%; Nano Banana 2 Flash 1K and 2K, both at 75 credits by August, from 1.4% to 26%. Within a model the prices moved little, and in both directions: the one clear cut went against the trend, Nano Banana 2 Flash 2K from 100 credits down to 75. Over the same 7 months people needed fewer generations per kept image. In the image generator the ratio fell every month, from 4.7 in February to 3.6 in August, down 23%. Sessions over 20 generations went from 12% to 8%, while the median session stayed at 4. The credits per kept image absorbed most of it: 278 in February, 302 in August, up 9% in the image generator and 19% across the tools that kept the same name all year. The try got much more expensive. The image you keep, much less.

Window

February 1 to August 31, 2026 (UTC), 7 complete calendar months. The window starts in February because the identifier that links a generation to its credit charge and its user exists from January 19, 2026. Monthly volume is not published as a trend: generations follow platform traffic, and this report measures ratios. The surface comparison uses the period from June 9, 2026, when API, MCP, and agent traffic can be attributed.

Population

Successful image generations on the platform’s image generation flows (Freepik and Magnific combined), all surfaces except internal traffic: 906,480,782 generations, the population of the two previous image reports restricted to the window. 852,592,410 of them, 94.1%, link to a credit record that carries the user, the tool, and the list-price credits; coverage rises from 92.8% in February to 99.35% in August. The generations that don’t link sit on the cheapest routes, so the cost figures are, if anything, understated in February and the 46% rise is a floor. Percentages about cost use the 829,144,478 linked generations with a list price; sessions use the linked generations; the download ratio uses the 821,287,756 linked generations made in the web app, 576,130,208 of them in the image generator.

Definitions

Credits are list-price credits: what a generation costs on a plan that pays per image. Generations on unlimited plans, 65% of priced generations, count at that list price even though the plan does not charge them one by one; their share fell from 78% in February to 46% in August. Free generations, 2.5% of the linked total and only from April, count as generations and as zero credits. Refunded generations, 0.3%, are left out of cost. Charges are per request; 98.5% of requests return one image, so credits per generation and credits per request are the same figure. A session is a run of generations by the same user in the same tool with less than 10 minutes between consecutive generations, dated by its first one.

A download is a download or copy event recorded for the same user, in the same tool, on the same day (UTC) as the generations; copies are 1.5% of the total. No record links a download to the image it belongs to, so this report never says which image was kept, only how many were generated for each one downloaded; both sides are restricted to the web app. Downloads on days without a generation in the same tool are not counted; 2 in 3 of them belong to tools outside this population. Adding every one of them would take the all-tools ratio from 4.6 to 3.8; the defensible range is 3.9 to 4.6. The per-user figure, 6 generations per download for the typical user, is the median over 1,085,259 users with at least 5 generations and 1 download, counted with downloads from every surface. Tools follow the credit record: the image generator, Spaces, Talk to image, and the editing tools. In July the platform renamed its retouch flow into Talk to image on both the generation and the download side, which breaks the all-tools download series; the monthly series is therefore shown for the image generator, whose share of days with a download is flat across the window. Tenure is the time since the user’s registration date on the day of the generation; 0.1% of user-days and 0.05% of sessions carry a registration date later than the generation and are excluded. Models carry their public names; the same model served through several routes is one row. Full definitions in the Report Glossary.

Limitations

1. The download ratio is an aggregate, not a per-image attribution

It counts downloads and generations by the same user in the same tool on the same day. It cannot say which generation was kept, and it counts an image downloaded twice as two downloads.

2. A session is a definition, not an observation of intent

Its median survives the 10-minute gap and the tool boundary; its mean does not, and rises from 8 to 12 generations in August with a 30-minute gap. Where a mean appears, its definition is next to it.

3. The interface generates in batches of 4

Each image in a batch is a separate request. A 4-generation session can be a single click.

4. Credits are list prices

On unlimited plans, 65% of priced generations, the marginal cost of a generation is zero, and those sessions average more generations at every price. The credits per kept image are what the same images would cost on a pay-per-image plan.

5. The rise in credits per generation is a change in which models people use, not in what models cost

At fixed prices the average rises as much as it did. Weighted by February's mix, prices within a model add 4 credits; weighted by August's, they subtract 5, because one model's price fell from 100 to 75. Both are small next to the 28-credit rise.

6. The editing tools are left out of the download ratio

They are 9.5% of generations, and their results are applied on the canvas and rarely produce a download event.

7. Sample-free, but not drift-free

The population moves by about 0.002% between runs as the data warehouse refreshes, which explains differences of a few hundred rows between tables.

Reproducibility

A second pass with its own queries recomputed the population and coverage, the credit distribution, the monthly means, the session statistics, the download ratios on a consistent base, the tenure buckets, and the price-by-plan table before publication, in two rounds: the first found 4 blocking issues and 11 corrections, among them the download base and the July renaming; the second verified the corrected figures and added 4 precision corrections.

Cite & data

Cite this report

Citation

Magnific (2026). It takes about 4 AI images to keep one. Creative Signals, September 2026. magnific.com/blog/cost-of-a-kept-image

Short form for press

“Across 906 million successful AI image generations on Magnific and Freepik between February and August 2026, people using the web app downloaded 1 image for every 4 they generated in the image generator, and 1 for every 4.6 across all image tools. A generation cost 75 credits at list price half the time, and the image kept about 300. Between February and August the average generation went from 60 to 88 credits, up 46%, as people moved to pricier models, while the generations per kept image fell from 4.7 to 3.6. The kept image went from 278 to 302 credits, up 9%, and up 19% across the tools that kept the same name all year.”