Nothing is free - navigating model cost

With a background in imagery, where even high-volume images required a physical set, products and humans, typing into the magic box with a prompt like “a fuzzy rabbit” and getting a photorealistic image of a, well... fuzzy rabbit in less than a minute was super impressive. Nothing like that existed a few years ago, there were no fuzzy rabbits made in a minute, and now it's limitless iterations. Where did this fuzzy rabbit come from? Did it exist as a real fuzzy rabbit someone photographed? Amongst the perplexity, this fuzzy rabbit's existence on my screen has flattened the value of an image.

So: limitless fuzzy rabbits, with only a subscription needed. Image value has flattened to the floor, and each generation feels like it costs nothing when your subscription gives you x amount of “credits” or “unlimited” generations. I wanted to know how many fuzzy rabbits I could get for my subscription, and I've tested a lot of platforms in the last year.

15,000 generations means I've generated an unbelievable number of bad images, even worse motion clips, sometimes over and over and over again, as I tried to get the model to just generate my fuzzy rabbit in a heroic pose.

60,000 credits sounds like a lot.

Almost every AI SaaS platform that offers the top image and video models has its own arbitrary system for charging usage. It's confusing, because bigger numbers feel like more value. And they are, until a single 10-second clip costs 5,000 credits and the 10,000 you had for the month is halved.

Take my usage in the screenshot below. At the time, I was on the Max plan with Krea: 60,000 credits a month for $100.

Most of it is image generation, averaging around 50 credits each, so hundreds of images across different models. However, the highlight here is the purple bar: Seedance 2, one of the best video models available at the time. That bar is a single 15-second, 720p generation, and it cost 3,600 credits.

As someone with no math ability, this is the math I use to convert credits to dollars:

1. Divide the generation cost by your allocation: 3,600 ÷ 60,000 = 6%
2. Apply that percentage to your plan price: 6% × $100 = $6

That one clip used 6% of my entire monthly allocation and cost $6 in real money. It's the same as 72 images.

And that assumes it works first time - it never does. Past the first few drafts, where you're fixing prompt structure and major issues, every attempt has to run at full spec, because lower-res drafts at 480p don't carry over, even when upscaled in post. If a usable clip takes three tries, that's 18% of the month and $18 for a single clip.

The levels of complicated

Almost all platforms, with a few small exceptions, call the APIs of the model owners. Krea receives your prompt for ChatGPT Image 2.5, sends it to OpenAI, gets charged, marks up that cost and returns the image to you. Strip away the UI, and all the platforms are the same. Here are the levels, from hardest and cheapest to easiest:

1 – Run the model via API > pure code, with instructions pointing to where your reference images are on your desktop. This is the cheapest, and the most difficult and unintuitive, but the cost comes direct from the model owner, like OpenAI, Google or Kling. It's technical rather than hard, but not a practical option for most people. You give up an easy time building workflows or keeping track of past generations this way.

2 – Run via a node-based tool > ComfyUI or RunComfy. These are node-based workflows, like using widgets and connecting dots depending on what the goal is; it could be as simple as the model's node and a text prompt. Frontier models aren't open weight, so ComfyUI reaches them through API nodes, charged per run. Nodes are more visual, with colours to help guide the learning curve, but again, no easy way to track generations over time, like a photo album.

3 – Run via a pay-as-you-go host > Replicate. Replicate is more forgiving. They charge you as you run the model, with their markup, so it's like option 1, but their UI is easy to follow and understand, and because the cost is in simple USD, it's easier to track. More functionality, but still not something easily trackable.

4 – Sign up for a subscription to a platform > Krea, Higgsfield, Figma Weave. Platforms here trade transparent pricing (USD per run) for the easiest and most feature-packed UI and experience. Backends here are much easier to follow visually and are set up for ease of use. Platforms like Krea have full node-based workflows, which are incredibly powerful for projects. Naturally there's a learning curve, as these platforms offer so many kinds of AI models: image, video, music, upscaling. However, at the time of writing, Krea has fully integrated Claude and ChatGPT into their platform, so with their Krea agent you should theoretically be able to bridge your vision and what the final set of images, videos or assets should look like.

As new models release, or when I'm researching platforms and their value, what I usually do is pick one model that's popular and stable. I find the API, or raw, cost listed by the model owner (OpenAI > ChatGPT Image 2.5). You can very easily find what OpenAI charges: look up “API cost [model]”, or ask ChatGeePeeTee, Claudio or Jimmenee. Then I find the model on the platform and set up my parameters. Most platforms will show you the generation cost, even without an account, which gives me whatever cost they've set for their platform. I take the subscription cost, convert it to the real cost, and measure that against the API cost. This is easier done with a chatbot.

It's also a good way to judge what a platform offers beyond generation cost: a polished, creative-friendly UI, transparent terms, or just stability, so you can still access your generations a year from now. Things like: are they going to move their app off the App Store, completely change their backend servers and where accounts point to, and break my Apple account sign-in on web? And now I'm speaking to an AI... about this AI platform, trying to regain access to my account. These would be good things to know, because no amount of cheap fuzzy rabbit generations is worth this level of headache.

Some of the landscape of AI slop

Models release often. Not in years, or even half-yearly; often it's months. Since mid-2025, I've seen this level of progression: Nano Banana was a breakthrough in mid-2025 > Nano Banana Pro was groundbreaking > ChatGPT Image 2 brought a level of realism and text/object accuracy I'd never seen outside real life (what is real life now?) > Seedream 5 Pro pushed a step ahead again, with its ability to generate some of the most aesthetically pleasing images > and now ChatGPT Image 2.5 Sunburst, which, in my limited testing, is the best model currently on the market and a very strong step above Image 2, which it succeeds. All in just over a year, and there are so many models that are intelligent and cost-efficient to use. These models rarely generate 6½ fingers on human subjects, whereas plenty of models before Nano Banana did. That's training data and optimisation from the megabrains building these models.

What I suspect often happens during model releases is that model holders like ByteDance or Google want to run a public beta test: release the model to as many people as possible by partnering with platforms with free usage for a period, and take in prompts and edge cases, where the model breaks, etc. That's valuable post-training data, and the easiest way to get it is to run free weekend or week-long generation promotions.

For a week in April 2026 and again in May 2026, Seedance 2 was unlimited on Krea. At $6 a run, I generated hundreds of videos, 99% of them slop. That's well over a thousand dollars' worth of generations for one user, and it wasn't out of the kindness of TokTik's heart. And Seedance 2.5 released at the end of July 2026.

What I'd suggest

What I have seen in the past year is platforms are moving into the enterprise space, which signals a solidifying and maturing product.

If you're an individual creator, know your landscape. Research platforms and their socials, find out how often they run promotions like this, and read comments from other users about any throttling or slow-lane generation queues that happen during those weekends. What I suspect is that platforms are given an allocation of free generations (the goal of platforms is to sell subscriptions, which is why free weekends are usually for higher-tier plans), their users overwhelm that allocation, and then any ability to stay on a schedule built around that week of free usage is broken. This was my experience with Seedance 2 in April, but mainly in May, where the throttling was very obvious.

If you're a business, timelines and deadlines matter more than cost, so much of this is not relevant. At production scale, AI pipelines have to align with traditional ones; I learnt this planning my own timelines on OK., knowing roughly what I needed to generate and which models best suited my vision. So the focus here would be to have clearly defined pools of what needs to be generated, add a buffer period for early testing and final polish (my 90% rule), and build a pipeline with a platform that focuses on ease of use and team-oriented features. Something where members can move in and out of projects, and all prompts and generations are tagged, retained and easy to navigate.

Here are some other things I've learnt that could be helpful

Free usage/credits don't usually grant commercial licence usage. On most platforms I've used, the lowest paid subscription is enough to grant commercial usage. Even if it's a monthly sub, as long as the generation was made and downloaded during the paid period, it should be valid for most commercial usage thereafter, but check each platform's terms. This is something I learnt the hard way with some SFX generations with ElevenLabs for Mirage.

Another important one is model training. Most casual or solo operators might not be aware that they're paying to use a model that is, in turn, using their prompts and outputs as useful data to train the next model. For a business with unreleased product, that potentially includes your product images. Business plans with some platforms have more robust enterprise terms than others. This is why you might see higher prices.

Knowing your landscape before committing time and production time to AI will help reduce situations where you’re reaching out to their support about their AI models, and hearing back from their AI.

*No fuzzy rabbits were harmed during the making of this article OR when generating image after image.