Not really worth it... yet.

I finally saw Adobe’s credit pricing after the unlimited use trial expired, and it explains a lot.

The issue isn’t image quality. Some of the partner models, looking at google's banana pro, can produce solid results when they work. The problem is the value model.

At 40 credits per generation, the basic plan caps out around 100 images per month. That could be reasonable if the system were reliable, or if there were any mechanism to recover credits lost to failed generations, hallucinations, or outright errors. There isn’t. OpenAI’s model at 60 credits per attempt only magnifies the problem.

In practice, you’re often spending a large fraction of your monthly credits just to get a single usable image. Once you hit submit, the credits are gone, regardless of output quality. That’s not a great value proposition for a probabilistic system. It’s especially hard to justify at Creative Cloud pricing.

During the recent “unlimited” promotion, which crashed frequently, I ran repeated tests to measure success rate. Google’s models landed usable results roughly 60 percent of the time overall. For simple, low-context tasks like background replacement, that climbed to around 75 percent.

I’m a machine learning engineer, and this wasn’t a casual test. It was an evaluation of whether these tools could add real value to my business and where they currently stand in terms of capability versus cost. There’s real promise here, but the economics don’t line up yet.

Adobe and its partners seem focused on model capability without equal attention to compute efficiency, reliability, and failure tolerance. In a credit-based system, those things matter just as much as raw output quality.

The message came through clearly.

You’re sorry.
Your servers are melting.