new flux.3 costs 2.8x more than seedream 5.0 flash. is it worth it?

flux.3 costs 2.8x more than seedream 5.0 flash for a stronger camera look. for high-volume work, seedream wins on price. we ran both on the same five prompts, one shot each, and below is everything: the prompts, the numbers and what broke along the way.

the setup

one prompt per task. one shot, no reruns, no edits, same aspect ratio on both sides.

• models: flux.3 image from black forest labs and seedream 5.0 flash from bytedance seed. both landed on openrouter on october 1.
• how we called them: both through openrouter's image api (/api/v1/images), with a small powershell script that sends the prompt, the aspect ratio and the resolution, and saves the png. one exception: seedream's plush toy is from our first round, sent through /chat/completions before we switched endpoints; it came back at the same 2048×2048.
• resolution: 2k on both sides. aspect ratio 4:5 for the mantis, tokyo and thermal portrait, 1:1 for the plush toy, 2:3 for the chrome head.
• prompts: four of the five were written from a reference image the editor sent: we described its subject, light, lens and texture, and fed the same text to both models. the plush toy prompt came from a plain brief: "super fluffy, minimalist".
• no cherry-picking: no best-of-n, no upscaling, no inpainting. what came back first is what you see.

the numbers

price per image

1. seedream 5.0 flash – $0.018
2. flux.3 – $0.05

the flux.3 price comes with a caveat. openrouter lists flux.3 at $0.10 per image at 2k, with a 50% launch discount, so we were billed $0.05. if the discount ends, the gap doubles to about 5.6x. seedream's $0.018 is its list price; the old chat endpoint didn't return a cost, the image api did.

1. mantis macro (4:5)

Extreme macro photograph of a green praying mantis head facing the camera, large translucent compound eyes with visible hexagonal facets, the whole head covered in hundreds of tiny crystal-clear dew droplets of different sizes, thin reddish-brown antennae beaded with water drops like a string of pearls, delicate mouthparts and palps with fine hairs, soft green body blurred below, pure black background, focus-stacked razor-sharp detail, soft diffused flash lighting with gentle highlights in each droplet, award-winning wildlife macro photography

both are close to a real photo. seedream went perfectly frontal and symmetric. flux.3 piled on more droplets, drew the mouthparts in finer detail and let the body run blurred below the head, as the prompt asked.

2. y2k tokyo fisheye snap (4:5)

Candid Y2K Tokyo street fashion snapshot shot with an ultra-wide fisheye lens and direct on-camera flash, two stylish young adults posing outside a brightly lit shop entrance with chandeliers and karaoke signs in Japanese (カラオケ), a young woman with curly bleached-blonde hair and a knitted flower headband, white crop top, low-rise baggy jeans with fringe belt, headphones around her neck, pouting and reaching toward the lens with both hands, a young man beside her with spiky black hair, small green wraparound sunglasses and a black t-shirt, his hand in the extreme foreground with chunky silver rings stretched toward the camera, exaggerated fisheye distortion, harajuku gyaru subculture, 2000s digital camera look, saturated colors, slight grain

this is where flux.3 earns the "camera look". it pushed the fisheye harder, added a round vignette and a wall of stickers around the doorway. seedream's shot is cleaner and reads like a real 2000s digicam photo, with カラオケ as a big readable neon sign. flux.3 got the big signs right too, but the small print around them is noise.

3. fluffy plush toy (1:1)

Product design render of a minimalist plush toy, a simple round blob-shaped creature covered in ultra-fluffy long pastel-pink fur, two tiny black bead eyes and no other features, sitting on a seamless soft beige background, studio softbox lighting, every fur strand visible, gentle shadow underneath, clean Scandinavian minimalism, premium toy catalog photo

the one task where flux.3 followed the brief more literally. it made a round pom-pom with two bead eyes and nothing else. seedream made a cuter, shop-ready character, but added a pear-shaped body, stubby arms and feet that the prompt ruled out with "no other features".

4. chrome art-toy head (2:3)

Studio product shot of a polished chrome metal head of a cute designer art-toy girl in collectible blind-box vinyl figure style, chibi proportions with a big round face, chubby cheeks, large glossy black almond eyes, tiny nose, small pouty lips, short bob haircut with blunt bangs and two little hair buns sculpted in liquid silver, small devil horns, engraved Japanese characters 東京 on the forehead, mirror-polished reflective silver surface with soft highlights, centered and floating on a pure black background, dramatic low-key lighting, high-end luxury toy photography, ultra-detailed 3D render

both wrote 東京 correctly, no broken strokes on either side. seedream made a mirror-chrome product shot with a soft shadow under the head. flux.3 went for a softer, satin metal that looks more like a toy than a luxury object.

5. thermal portrait (4:5)

Surreal portrait of an androgynous figure with head slightly bowed, eyes looking down, wild spiky hair radiating outward like sharp rays, rendered in a soft-focus solarized thermal-camera aesthetic, inverted colors: face in deep cobalt and navy blue shadows, glowing thin orange slits for eyes under sparkling frosted brows, hair and shoulders blooming from hot pink and coral red into blown-out white, heavy airbrush grain and stippled noise texture, dreamy blur and halation, dark midnight navy background, Y2K experimental album cover art, analog film scan look

seedream stuck closer to the reference. it kept the exact palette: cobalt face, pink hair blooming into white, thin orange eye slits, a halo behind the head. flux.3 took the same prompt darker, toward a film-negative look, with the eyes closed and the lashes glowing orange.

what broke along the way

the test took one session and several fixes. if you're wiring either model into a pipeline, these are the ones that will bite.

1. flux.3 lives on a different endpoint. our script started on /chat/completions, which works for seedream. flux.3 only answers on openrouter's dedicated image api, /api/v1/images, so we moved the whole script there. it takes prompt, aspect_ratio, resolution and up to 10 input_references, and returns the image as base64 plus the billed cost.
2. quotes broke a prompt. a prompt with text in double quotes (a poster with "URBAN" and "LOST IN THE NOISE") split on the command line and failed with "a positional parameter cannot be found that accepts argument 'THE'". we switched those quotes to single ones and, for good, moved every prompt into a utf-8 text file the script reads with -PromptFile.
3. non-latin text in prompts. 東京 and カラオケ risk getting mangled when passed through the windows command line. reading prompts from utf-8 files fixed that too, and both models rendered the characters correctly.
4. the task set changed twice. round one was a jumping spider macro, a "metalheart" magazine cover, the plush toy, an anime poster and an image edit. we then rewrote the tasks from reference images: the mantis replaced the spider, a tokyo street snap replaced a poster test, a chrome art-toy head replaced the anime poster. the edit test and a green hologram city were cut. the plush toy is the only round-one task left.
5. 6–8 mb pngs broke our video. we recorded the side-by-side clip by screenshotting the comparison page in headless chrome and stitching frames with ffmpeg. flux.3's mantis is a 6 mb png, and chrome took the screenshot before it loaded, so the right half of the first slide was a grey box. we raised the wait time and added a check that re-shoots any frame with a grey placeholder, then pulled one frame per pair out of the final mp4 to confirm all ten images were there.
6. ffmpeg wouldn't unpack. windows' 260-character path limit stopped Expand-Archive in our deep temp folder. we pulled ffmpeg.exe straight out of the zip into a short path instead.

the takeaway

flux.3 costs 2.8x more for a stronger camera look. for high-volume work, seedream wins on price.

sources

• FLUX.3 Image on OpenRouter – $0.10 per 2k image, 50% launch discount at the time of the test
• Seedream 5.0 Flash on OpenRouter – $0.018 per image
• FLUX 3 announcement – Black Forest Labs, jul 23
• OpenRouter image generation docs – the /api/v1/images endpoint we used