Freepik AI Image Generation API: Pricing and Magnific Upscaler

Freepik's developer API did not stay Freepik for long. Since late April 2026 the whole platform, stock library and AI tools included, sells under the Magnific name, and the old documentation address now redirects to docs.magnific.com. If you are budgeting an AI image generation API for a product, the rename matters less than the price sheet underneath it, because your bill has two very different parts: a cheap image generator (Mystic plus a long list of Flux and Seedream models) and an upscaler whose price follows the output size in pixels.

This article lays out what the docs state, what third-party trackers report, and where the two disagree, so you can estimate a real monthly cost before writing a line of integration code. Anything marked documented comes from the official docs. Anything marked reported comes from outside sources and deserves a check against your own dashboard.

Freepik Is Now Magnific

Several 2026 pricing write-ups date the switch to April 28, 2026, when Magnific became the single brand for the stock assets and for an AI suite spanning image, video, audio, 3D and design. For API users, the visible effects are small but real.

What the Rebrand Changes

• Docs address: docs.freepik.com answers with a permanent redirect to docs.magnific.com (documented).
• Authentication header: the current docs name the header after Magnific instead of Freepik (documented).
• Old hosts: one write-up updated in late September 2026 says both api.freepik.com and api.magnific.com were still accepted during the migration (reported).

💡 Tip: Put the host and header names in one config file. The next rename then becomes a one-line change instead of a search through your codebase.

Endpoints You Will Call

Three endpoints do nearly all the work for image projects. Every one of them is asynchronous: you submit a job, receive a task_id and a status, then either poll for the result or let a webhook_url call you back. Mystic tasks move from CREATED to IN_PROGRESS and end either in a finished state or in FAILED (documented).

A typical integration follows the same four steps for all three:

1. Send the request with your prompt or source image URL and, ideally, a webhook_url.
2. Store the returned task_id next to your own record, such as an order or a product ID.
3. Wait for the webhook, or poll the task endpoint at a modest interval.
4. Download the output right away and save it to your own bucket.

How the API Bills You

The first thing to settle is which balance pays for your calls. Mixing it up is an easy budgeting mistake with this API.

Credits, Not Subscriptions

The docs state that every AI request, whether image, video, audio, upscaling or editing, deducts credits from your organization's balance (documented). Consumer plans are a different product. Trackers list Premium at €12 per month on annual billing with 240,000 credits a year, and Premium+ at €27 per month with 600,000 credits (reported). Several write-ups describe the API as a separate, pre-paid developer service, so confirm which balance your requests draw from before you ship. A cheap way to do it: run one 1K Mystic generation, then compare the balance before and after. That single test gives you your own per-image price, measured on your own account.

Per-Image Prices From Trackers

The official docs publish an upscaler price table but no per-model generation table. The numbers below come from third-party price trackers, so treat them as a starting estimate.

⚠️ Watch for outliers: one write-up updated in late September 2026 still lists Mystic generation at 0.40 per 100 creations, about 0.004 each. That is roughly ten times below the Flux 2 Pro price above. A gap this size is a good reason to read the live price list before you budget.

Free Credit and Spending Cap

Write-ups report roughly 5 in free starter credit (euros in one, dollars in another) and a self-serve ceiling of 500 per month before you have to talk to sales (reported). If your plan needs more than that, ask for a custom agreement early, because the same pages say higher limits go through the sales team.

Mystic and the Model Catalog

Mystic is Freepik's in-house text to image engine, and its parameter list is long enough to deserve its own table.

Mystic Parameters Worth Knowing

Two details save debugging time. First, prompts accept character references written as @character_name or @character_name::strength. Second, filter_nsfw cannot be switched off for standard users, so adult content will not pass through this API at all. By name, super_real and editorial_portraits point at photographic work, while zen and fluid sound closer to painterly output. Run a small test batch of each before committing to one, using the same prompt each time, so the only variable is the model.

Other Text to Image Models

Beyond Mystic, the docs list Flux Kontext Pro, Flux 2 Pro, Flux 2 Turbo, Flux 2 Klein, Flux Pro 1.1, Flux Dev, Hyperflux, Seedream 4, Seedream 4.5, Z-Image Turbo and a Runway text to image endpoint (documented). That breadth is the real selling point: one integration and one balance, many generators.

The names hint at the intended jobs. Kontext Pro belongs to the Flux Kontext line, which is built around editing an existing picture from a text instruction. Turbo and Klein variants suggest speed and lower cost per image. Seedream 4.5 and Flux 2 Pro are common picks for polished marketing visuals. Because the catalog moves quickly, read the live model list before you hard-code a model name, and keep the name in configuration rather than in your source.

Magnific Upscaler Prices by Size

Upscaling is where the bill grows. Both modes support 2x, 4x, 8x and 16x factors, and both are priced by the pixel area of the output, not by the number of calls (documented).

Creative Mode

The creative endpoint accepts scale_factor, optimized_for, prompt, creativity, hdr, resemblance, fractality, engine and webhook_url. In practice it behaves like a generator steered by your photo: it adds texture and depth that were never in the source. That suits illustration, concept art and soft photos that need a lift. It is a poor fit for faces you must keep faithful or any image with text, because invented detail can alter both.

The docs advise sending the original image by URL whenever possible. Base64 uploads and canvas exports lose quality before the upscaler ever sees them.

Precision Mode

Precision mode, introduced in mid-2025 according to reviews, enlarges the picture without hallucinating new content. It is the safer choice for logos, interface screenshots, product shots and scans. The docs recommend an input of at least 256 to 512 pixels on the shortest edge, PNG for graphics, high-quality JPEG or PNG for photos, and moderate factors of 2x to 4x, since pushing further tends to amplify noise.

A quick way to choose between the two modes:

What Three Price Points Reveal

The docs give three worked examples, identical for both modes. The megapixel columns are my addition, to show the pattern.

From only three points, two things look likely (my inference, not a published formula). Small outputs seem to sit on a floor near ten cents, and large outputs settle around two and a half cents per megapixel. The practical lesson: a tiny image upscaled 8x costs more than a full HD frame upscaled 2x, so decide the final pixel size before you pick a factor.

Rate Limits and Gotchas

Request Ceilings and Expiring Links

The documented limits are generous for most projects (documented):

• Per IP address: 50 hits per second over a 5-second window, and an average of 10 hits per second over 2 minutes.
• Per credential: 30,000 requests per minute.
• Daily limits: none for most endpoints. A few helper services, such as image to prompt and search, cap at 100,000 requests per day.
• Custom limits: available through the sales team.

The per-IP numbers matter if a worker fleet shares one outbound address. Ten workers that each fire five requests per second will far exceed the 2-minute average of 10 per second. Add a small queue and an exponential backoff when a request is throttled, and the limit becomes invisible to your users.

The limit that actually bites is not a rate at all. Output links expire 30 minutes after generation, so copy every result to your own storage the moment a task reports that it has finished. Prefer webhooks over tight polling loops, and log every FAILED task so you can retry it before the window closes.

Budget Math for Real Projects

A 1,000 Image Catalog

Say an online shop wants 1,000 product visuals, each generated once and upscaled to 4K. Using the reported Flux 2 Pro price and the documented upscaler price:

Upscaling costs roughly five times more than generating, so the cheapest saving is curation: pick the winners first and upscale only those. Small thumbnails behave the same way. Upscaling 5,000 images from 640×480 to 1280×960 at €0.10 each comes to €500, which equals the self-serve monthly ceiling reported earlier.

💡 Tip: Generate at the resolution you need and skip the upscaler when 2k output already fits the layout. Every upscale you avoid is a full unit price saved.

Five habits keep a bill flat as volume grows:

1. Upscale approved images only. Never run the upscaler on a draft.
2. Start at 2x. Move to 4x or 8x only when a real layout demands the pixels.
3. Log every task. Store the task ID, the model, the output size and the cost per project, so a surprise on the invoice can be traced in minutes.
4. Add a 10 to 15 percent buffer for re-runs when a result misses the brief.
5. Save outputs on arrival. A link that expires after 30 minutes is a re-run you pay for twice.

Picasso IA Models That Fit

If you want to compare quality before committing to any API, Picasso IA hosts close alternatives for both halves of the workflow, and you can test them in a browser first.

Text to Image Equivalents

• Flux 2 Pro and Flux 2 Flex for the same Flux family that Magnific lists in its catalog.
• Seedream 4.5 to compare against the Seedream 4.5 endpoint.
• Nano Banana Pro and Imagen 4 as general purpose alternatives from Google.
• P Image when you want fast drafts at volume.

Upscaler Equivalents

The pairing is one of spirit, not identity. Each model has its own look, so run the same test photo through two or three before you judge.

How to Use Clarity Pro Upscaler

Clarity Pro Upscaler is the closest match to Magnific's creative mode because of its single creativity control. It keeps facial structure, skin pores and hair strands intact while it enlarges.

1. Open the model page and upload the image. The image is the only required input.
2. Set scale_factor. The default is 2. Choose 2, 4, 8 or 16 depending on your target size.
3. Tune creativity. The default is 0. Negative values stay close to the original, positive values add texture and depth.
4. Pick output_format. PNG is lossless and is the default. JPG uses 95% quality and gives smaller files.
5. Run the model and download the result. Save it straight away.

💡 Tip: Start at 2x with creativity at 0, then raise one setting at a time. Changing scale and creativity together makes it hard to see which one caused a change.

Try It on Picasso IA

Developers can also call Picasso IA programmatically. Its API sits at https://api.picassoia.com/v1 and follows the same pattern as the Freepik flow above: create a prediction, poll it, fetch the result. When this article was written it offered four models (image, image editing, video and video with audio) with 5 concurrent predictions per account. The site's wording on cost and plan requirements was not consistent at that time, so confirm it on the pricing page before you build on it.

For a quick test with no integration at all, take a soft, low-resolution photo, run it through Clarity Pro Upscaler at 2x, then try Topaz Image Upscale on the same file. Generate a fresh image with Flux 2 Pro and send it through the same steps. Within an afternoon you will know which look fits your project, and you will have real numbers to compare against the Magnific price list. Pick a photo, open a model and run your first test today.