Hi everyone, I’ve been using DALL-E 3 through the API for a couple of months now as part of a vendor evaluation for our content team’s needs. I’m deep in the RFP process for an image generation solution, so I’m logging every detail—latency, output consistency, TCO implications for high volume, you name it.
The documentation and the dashboard both consistently mention a typical generation time of about 15 seconds. But in my controlled tests over the last two weeks, I’m seeing significantly higher averages, and it’s throwing off my total cost of ownership calculations. My process has been to generate batches of 10 images per prompt (standard quality, 1024x1024, no variations) and measure the response time.
Here’s a snapshot of my last batch from yesterday:
* Prompt: "A detailed technical illustration of a cloud server stack, isometric view"
* Generation times recorded: 22s, 19s, 34s, 28s, 18s, 41s, 23s, 26s, 31s, 20s
* Average for this batch: 26.2 seconds
* Longest wait: 41 seconds
That’s a pretty far cry from 15 seconds. I’ve controlled for what I can think of:
* Running tests during off-peak hours (late evening EST).
* Using a consistent, low-latency connection.
* No parallel requests that would queue.
* Prompts are descriptive but within a similar character range.
This isn’t a one-off. The pattern holds across hundreds of requests in my logs. The variability is also a concern—a 34-second difference between the fastest and slowest in a single batch is substantial when you’re building workflows.
My core question is: is this the expected real-world performance others are experiencing? The 15-second figure seems more like a best-case scenario rather than a median. For those of you also evaluating for professional or procurement purposes:
* Are your generation times consistently around the 15-second mark?
* Have you found any parameters (prompt length, aspect ratio, quality setting) that reliably impact speed?
* Does the platform (API vs. ChatGPT Plus vs. Microsoft Designer) make a significant difference in latency?
I’m trying to benchmark this against other vendors, and latency directly affects user adoption projections and infrastructure planning. If the average is actually mid-to-high 20s, that changes the feasibility for certain real-time use cases we’re considering. Any data or experiences you can share would be incredibly valuable for my comparison matrix.
Of course you're seeing longer times. Their "typical" generation time is a best-case scenario, probably calculated under laboratory conditions with the simplest possible prompts. The moment you add any complexity, like "detailed technical illustration" and "isometric view," you're adding computational steps they don't factor into their marketing material.
This is a classic vendor move. They sell you on a low, efficient-sounding metric to get you hooked on the cost-per-call math, then the real-world latency blows up your operational timelines. Have you factored in how these variable times affect your team's workflow efficiency, not just the raw API cost? A 41-second wait for a single asset can cascade through a production queue.
You mentioned you're in an RFP process. Make sure you're baking these real-world averages, including the outliers, into your SLA requirements. If they won't guarantee a percentile bound on latency, their "15 seconds" is just a suggestion.
Trust but verify.