I Ran 6,000 Product Images Through Doubao Vision Pro. Total Bill: $5.
**AI Export Lab** | July 2026...
**AI Export Lab** | July 2026
Numbers first.
- **Images processed:** ~6,000
- **Model:** Doubao Vision Pro 1.5
- **Total spend:** about $5 US
- **Per-image cost:** roughly $0.0008
This is not a marketing number. This is my actual bill from a batch job on the `jjradiator.com` product catalog. I re-ran the totals before writing this post to make sure the number is correct.
If you have a project that's been stuck because "the images will cost too much to process" — read on, because that reason expired sometime in 2024 and nobody sent out a memo.
---
What $0.0008 Per Image Actually Buys
For each image, the model returned:
- All visible text — part numbers, OEM codes, dimensions, brand names
- A structured description of what the image contains
- Optional: extraction of specific fields I asked for by name (e.g., "return only the OEM number and the width in mm")
The output was clean enough to feed directly into a downstream text AI without a cleaning pass. When it was wrong, it was **consistently** wrong in a way I could fix at the pipeline level — one prompt tweak, one re-run — instead of row by row.
That's the metric that matters. Not raw accuracy. Predictable failure modes.
---
The Comparison That Made Me Do This
To put $5 into context, here's what the same job would have cost through other channels I actually looked at:
- **Data entry firm (Philippines).** Quoted $1,400. Turnaround: 2-3 weeks. This was the cheapest human option I found.
- **Data entry firm (China).** Quoted ¥8,500 (~$1,200). Turnaround: 10 days. Slightly cheaper because they were nearby, still four figures.
- **In-house intern.** $10/hour × 30 seconds/image × 6,000 images = ~$500 in labor. Plus my time to train them, which I priced at zero out of denial.
- **DIY manual, me.** 30 seconds per image × 6,000 = 50 hours of my life. Not happening.
- **Doubao Vision Pro batch.** $5. Overnight run.
I want to be careful here — data entry firms are still the right answer for jobs that need judgment, bilingual review, or handling ambiguous cases a customer will complain about. For pure "read what's on this image and give me the text," the economics have completely inverted.
---
The Free Tier Is How You Start
Most major vision APIs give you a free monthly quota when you sign up. For most personal projects — a few hundred images per month — you never leave that tier.
In my case that free tier is Doubao Vision Pro through Volcengine Ark. Mainland China only, requires a Chinese phone number. If you're outside China, the equivalent shape is easy to find:
- **Qwen-VL** on Alibaba Cloud International — English dashboard, foreign card sign-up, free monthly quota on new accounts
- **DeepSeek** international API — global sign-up, cheap tokens (vision support more limited than dedicated providers, but improving)
- **Xiaomi's overseas AI platform** — Google account sign-up
- **OpenRouter** — one account, dozens of models including Chinese open-source ones, no Chinese phone number
- **Gemini, Claude, GPT-4o** — vision built in; more expensive per image but zero setup friction
Pick the one you can sign up for. Each has a free tier of some kind.
This is what actually changed for me:
**Old pattern:**
- Get an idea
- Estimate cost
- Decide it's too expensive
- Do nothing
**New pattern:**
- Get an idea
- Register for the free tier of whichever vision API works in your country
- Run a small pilot inside the free quota
- See if the output is worth expanding
- If yes, upgrade. If no, walk away.
You verify before you spend. Nobody has to approve a budget.
---
When You Cross Into Paid Territory
Rough rule of thumb from my usage:
- **Under 500 images/month** — free tier covers you.
- **500-5,000 images/month** — you'll spend a few dollars.
- **5,000-50,000 images/month** — still cheap in absolute terms, but worth a small monitoring script so you notice weird spikes.
- **Above 50,000 images/month** — talk to the provider about volume pricing.
Put that alongside your other business inputs — hosting, domains, one Zoom subscription, one Slack seat — and image processing is one of the cheapest line items you have.
---
The Five Things That Kept My 6,000-Image Bill At $5
Nothing exotic. Just discipline.
**1. Batch, don't stream.**
Send images in batches instead of one at a time. Fewer API calls means less overhead. Most modern vision APIs accept batches directly.
**2. Ask for exactly what you need.**
The more specific your prompt, the fewer tokens in the response. *"List the OEM number and dimensions in JSON"* costs less than *"describe this image."* By a lot.
**3. Cache aggressively.**
Store every result on disk keyed by the image hash. If your pipeline might re-run — and it will — a re-run on cached images costs $0. I've re-run one of my catalogs four times over the past six months without paying twice.
**4. Use the smallest model that works.**
Doubao Vision Pro 1.5 is not the biggest vision model available. It's the one that hits my quality bar for OCR-style tasks. If your job needs a bigger model, spend more. For most catalog work, small is enough.
**5. Retry on failure, don't restart.**
Save your progress. If image 3,847 fails, the retry should only re-run image 3,847, not images 1 through 3,846. Sounds obvious. First time I ran a big batch, I did not do this. Cost me an extra $2 and 90 minutes.
None of this is clever. It's just the difference between $5 and $50 for the same job.
---
What You'll Still Get Wrong
Being honest about accuracy so nobody hits this and blames me:
- **Handwritten text** — vision AI is good but not perfect. Assume 5-15% error rate on handwriting. Budget human review.
- **Overlapping text on graphics** — part numbers stamped over a drawing of the part will sometimes come back with characters swapped.
- **Low-contrast scans** — old photocopies of paper catalogs are still painful. Better than they were, still not clean.
- **Small text in a big image** — if the OEM number is 2% of the image area, some vision models will miss it. Crop first or use a model that handles resolution well.
I lost half a Sunday to the "small text on big image" one before I realized what was happening. Now I pre-crop where I can.
---
The Bigger Picture
For years, "we can't do that project because the data is trapped in images" was a real reason.
It isn't anymore.
If your business has boxes of paper catalogs, or a shared drive full of supplier photos, or years of scanned invoices — that's not dead weight. That's unindexed inventory nobody has been able to afford to process.
For $5 and one weekend, you can find out what's in there.
Very small bet. Potentially large payoff.
---
**What's sitting on your hard drive?** If you have a specific pile of images or PDFs you've been meaning to process, drop me a note on the [contact page](/contact/). I'm building a list of "unusual things people have unlocked with cheap vision AI" and yours might be one.
*This post is part of the cornerstone series [How I Turned a 600-Page PDF Into a Product Database for Under $10](/ai-export-lab/how-i-turned-a-600-page-pdf-into-a-product-database-for-under-ten-dollars/). See also [How to Give DeepSeek Eyes](/ai-export-lab/how-to-give-deepseek-eyes-vision-api/).*