Seedream 4.0 vs 4.5 vs 5.0 Lite on Sume: price, ratios, model ids

Sume lists three Seedream models, and the cheapest is the oldest. Seedream 4.0 bills $0.0325 per image, Seedream 5.0 Lite $0.04375, and Seedream 4.5 $0.05, each computed as the list rate times Sume's 1.25 margin. All three take reference images and return up to four images per call.

Side by side

Ids to watch

Public ids are org/slug. The legacy bare ids still work as aliases, and they are not all the obvious slug: seedream-v4 is Seedream 4.0, and seedream-4-5-edit is Seedream 4.5. If you pinned the bare id in old code, it keeps working; for new code use the bytedance-seed/... form from the Image API page.

What is the same

• Ratios: 1:1, 16:9, 9:16, 4:3, 3:4, 4:5, 5:4, 3:2 and 2:3. That covers an Instagram 4:5 feed image and a 9:16 vertical cover.
• Edit support: all three accept input_references, up to 10 per request on the Images API.
• Output: png, jpeg or webp, and n up to 4 per call in the catalog.
• Custom pixels through image_size (Seedream is one of the families that takes them).

What differs

Only Seedream 4.0 lists auto as a ratio, which is handy for edits that should keep the source shape. The price spread is small in absolute terms: 100 images cost $3.25 on 4.0, $4.375 on 5.0 Lite and $5.00 on 4.5, which are straight multiples of the per-image rates above.

Sume does not publish a quality ranking between the three, so choose by running your own prompt set. A 20-prompt comparison costs under a dollar on any of them: 20 x $0.05 = $1.00 at the top rate of the three, and less at the others.

A three-model test plan

Take ten prompts that look like your real work: two product shots, two portraits, two text-in-scene cases, two edits from a photo, and two wide scenes. Run each prompt once on each of the three ids with the same aspect_ratio. That is thirty images, and it costs 10 x ($0.0325 + $0.04375 + $0.05) = $1.2625 billed.

Score the results blind: hide the model name, look at crops at the size you deliver, and write down what failed (hands, text, product edges). Then repeat the edit cases with input_references to see which model keeps the subject. Pin the winner for each job type, not one model for everything, since the price gap between rows is small enough that routing by job type is cheap to run.

How to choose

Pick 4.0 when cost matters most and auto ratio helps. Pick 5.0 Lite when you want the newer family at a lower price than 4.5. Keep 4.5 when your prompts are tuned to it. Test with n: 4 and compare crops at your real delivery size, then pin the id. Pricing and ratios are read from the catalog, so a recheck of GET /v1/images/models/{id}/endpoints before a big batch is cheap.