Retouch 200 product photos: Image API, Agent Completion or Format?
For 200 identical retouches use POST /v1/images: fixed model, price per image, plan queue limits. Use an Agent Completion only if each photo needs judgement.

For 200 photos that all need the same retouch, call POST /v1/images 200 times. You choose the model, the prompt and the quality, the price is per image, and the queue limits for your plan are documented. An Agent Completion or a Format run earns its extra cost only when each photo needs a decision, such as "fix the dust but keep the scratches that are part of the product."
The three routes side by side
All three accept images, and all three are on the same Sume key. The differences are in who decides and how you pay.
| Route | Who picks the model | Result | Cost control | Input limit |
|---|---|---|---|---|
| POST /v1/images | You | data[].url, sync up to 30 s, else 202 | price per image by quality | 0 to 16 references by model |
| POST /v1/agent/completions | The agent | output.images after polling | generation_spend_cap_usd, required | 30 images, 30 MB each, 500 MB per run |
| Format run | The attached Format | run output after polling | run limits in the Formats docs | same attachment limits as completions |
Throughput is a plan limit, not a route choice
Generation admission applies by plan: Free processes 1 job and accepts 6, Pro processes 4 and accepts 24, Startup 8 and 48, Scale 20 and 120 (Generation admission). So 200 photos are always a feed-the-queue problem. Use a worker pool the size of the processing number and retry on 429.
Price check on the cheap route
Ideogram 4.5 at quality: low is $0.0375 billed per image on Sume (fal list $0.03 x 1.25, read 2026-10-05), so 200 low-quality passes are $7.50 and 200 at the default medium are $15.00. Use low to proof a sample of ten before you spend the other 190 at the quality you want to ship.
Pick by the job
- Same fix on every photo, same shape out: Image API, no aspect ratio, so every output keeps its source shape.
- A written brief per photo that changes: Agent Completion with attachments and a spend cap per run.
- A repeated workflow someone else will run again, with its own rules: a Format run.
- Fewer than 20 photos once: whichever you can already call. The difference is minutes, not money.
Limits
A cheap route is only cheap if the first pass is right. Retouch is the case where a model can change a logo, so sample, diff against the source, and keep a person in the loop for products with printed text. An agent will not make that less true, and it makes the cost per photo harder to predict.
Sources
Related posts
More in Comparisons
- Rime Mist v3 vs Coda vs Sume TTS: cost for 1M characters
Rime lists Mist v3 at $0.03 and Coda at $0.05 per 1K characters. Sume TTS 1.0 is $0.0475. Here is the cost of 1M characters and a 20,000-character call.
- Runway: no RPM limit but a daily cap; Sume answers 429 two ways
Runway documents no per-minute limit but a rolling 24-hour generation cap. Sume returns 429 queue_full or rate_limited, with ratelimit headers.
- Runway dev MCP login needs Claude Code 2.1.186+; Sume: mcp login
Runway's dev MCP logs in with OAuth in the terminal on Claude Code 2.1.186+, --no-browser from 2.1.191. Sume uses the same two commands at mcp.sume.com.
- Runway SDK needs Node 18 or Python 3.8; Sume SDK is Node 18, no deps
Runway lists Node 18+ and Python 3.8+. Sume's TypeScript SDK 0.2.0 runs on Node 18+, Bun, Deno or Workers, no runtime deps; no Python SDK is documented.
Written by Sume