We run a small performance marketing agency. Switched our image generation mostly to Leonardo AI about 4 months ago. I've been tracking our costs and output to see if the hype is real for a business use case.
Here's our typical monthly breakdown on the **Scale plan ($48/month)**:
* **Spend:** $48 base + ~$120 in top-ups for extra tokens. So **~$168 total**.
* **Output:** Roughly **2,300 - 2,700 images** per month. That's a mix of ad creatives, blog graphics, and product mockups.
* **Key workflows:** Using Alchemy, Prompt Magic, and training a few PhotoReal LoRAs for client branding.
**The good:**
* Cost per image is **incredibly low** compared to some competitors we've tried (looking at you, Midjourney). Comes out to about **$0.06 - $0.07 per image** for us.
* The **API access** (included in Scale) lets us integrate into some of our internal tools. Big time-saver.
* Fast generation speed means we can iterate on client feedback live in meetings.
**The not-so-good:**
* The **token system** feels a bit opaque. Some complex prompts with Alchemy V2 eat credits way faster.
* We **have** to top-up. The base plan tokens are gone in the first 10-12 days. Wish there was a higher base tier between Scale and Enterprise.
* For hyper-realistic product shots, we still occasionally use another tool. Leonardo's PhotoReal is great, but sometimes the lighting/reflections need more tweaking than it's worth.
Overall, it's a clear win for volume and cost. The main hurdle is managing the token spend and understanding what drives it. Curious if other agencies have similar numbers or different workflows to stretch tokens further.
Trial number 47 this year.
Great to see another agency tracking this. I'm a technical lead at a mid-sized e-commerce brand, managing our own image generation workflows for product marketing and ads. We've been heavy Leonardo users for about 9 months, and I've also stress-tested the API for batch jobs.
Since you're on the Scale plan already, here's a breakdown from someone who's run it in production:
**True Cost Per Asset:** Your $0.06/image is in the right ballpark for bulk output, but that can triple for high-res, complex prompts using Alchemy V2 and upscalers. Our most expensive single brand-safe images (PhotoReal, hi-res, specific poses) can hit $0.18-$0.22 each when we factor in all the prompt iterations. The base plan tokens are really just an entry fee.
**API Stability & Speed:** For batch generating 500+ images overnight via the API, Leonardo has been solid. Latency is consistent, unlike some competitors where we'd see random timeouts. The queue system is predictable. We've never had an API outage that disrupted our morning routines.
**Opaque Credit Burn:** You nailed it. The biggest hidden cost isn't the top-ups, it's the experimentation. Training a custom model (even a LoRA) burns credits fast for dataset prep and training runs, and that's not always clear upfront. We budget an extra $50/month just for R&D on new model styles.
**Comparison to Direct Cloud:** We ran a parallel test on Amazon SageMaker with Stable Diffusion for two weeks. The cloud cost was slightly lower per image at massive scale, but the engineering overhead for model management, safety filters, and uptime wasn't worth it for a team under 50 people. Leonardo's all-in-one platform saves us about 10-15 engineering hours a month.
Given your agency's volume and need for client-specific branding (those PhotoReal LoRAs), I'd stick with Leonardo. It's the right fit for an SMB/mid-market team that needs reliable output without a dedicated ML engineer. If your top-ups are consistently hitting $120, look at the Enterprise tier - it might offer better bulk rates.
To make it clean, tell us: What's your ratio of simple blog graphics vs. complex, client-branded ad creatives? And do you ever hit the daily generation limits?
cost first, then scale
I completely agree about the opaque credit burn during experimentation. That's been our single biggest budget surprise, too. It's not just training models; it's the iterative process of getting a LoRA or style to actually work. We once burned through a $50 top-up just dialing in the prompt weights for a client's specific product texture.
Your point on API stability is interesting. We haven't done huge batch jobs yet, but the predictability is key for scheduling work. Have you found the need to implement any specific retry logic or error handling in your batch scripts, or does it just churn through the queue without intervention?
~jason