Splotch · image-model bake-off — 2026-09-10T21-10-52-556Z-image25-winning-prompt

Flare low is the leading challenger

A modest speed improvement is worth validating. The screen does not establish a universal visual winner or intrinsic cost savings.

Same winning prompt, fresh baseline

Fourteen synthetic drawings, one per category, two samples per configuration: 140 finished comparison images. Every call uses the shipped “Paint directly over” prompt (ADR-0118), with GPT-5.6 Sol at medium reasoning and the existing safety instruction.

What the averages say

Flare low: 23.0 s median versus 24.5 s for the baseline; p90 24.0 s versus 26.7 s. The broad run allowed three calls in flight. Mean layout scores: GPT Image 2 low 83.0; Flare low 85.5; Flare medium 86.8; Sunburst low 84.5; Sunburst medium 85.1.

The cost difference is mostly cache

Recorded mean cost is 3.16¢ for Flare low versus 3.53¢ for GPT Image 2 low. Removing the orchestrator cache discount puts every low tier near 3.94¢; the image tool itself averages about 1.52¢ on all three low-tier models. Do not interpret this run as a cheaper image-token price.

Look at the tradeoffs

The apple shows better blending in some medium-tier images. The colored balloon shows newer models sometimes preserving the red mark as a patch instead of coloring the balloon. The black scribbles show color changes despite high layout scores. Higher alignment scores do not automatically mean a better illustration.

Reliability and scope

The first pass produced 139 images and one baseline API 500. One targeted retry filled the missing image; the original failure is preserved in the run evidence. There were no refusals. The dark-paper row uses the base prompt without the Night Mode suffix. The blocked-content safety corpus and style suffixes were not tested.

Sequential confirmation

The separate counterbalanced check made 16 sequential calls over four inputs (eight per model). GPT Image 2 low: 24.2 s median, 31.8 s p90. Flare low: 21.8 s median, 23.0 s p90. Model order alternated AB/BA within each input and which model started alternated between inputs. This is a small confirmation sample, not a statistically powered latency study.

Before a production switch, validate the finalist with the actual style/night requests and the full safety suite. The measured visual differences do not justify skipping those checks.

Cost and speed

one row per candidate

Cost is what one generated image costs at list prices. Time is how long each call took, from request to response.

Sort

2 · lowprevious prod

3.5¢per image

$35 per 1,000 · 1.1× cheapest

24 smedian

p90 27 s · range 18–35 s

28 / 28images
flare · lowcurrent prod in production now

3.2¢per image

$32 per 1,000 · 1.0× cheapest

23 smedian

p90 24 s · range 17–27 s

28 / 28images
flare · medopenai candidate

3.7¢per image

$37 per 1,000 · 1.2× cheapest

24 smedian

p90 26 s · range 18–30 s

28 / 28images
sunburst · lowopenai candidate

3.2¢per image

$32 per 1,000 · cheapest

26 smedian

p90 29 s · range 19–31 s

28 / 28images
sunburst · medopenai candidate

3.7¢per image

$37 per 1,000 · 1.2× cheapest

28 smedian

p90 31 s · range 21–32 s

28 / 28images

medianp90 (9 in 10 calls finish by here)24 s, the limit of a single synchronous request

The dashed line is the app's original one-request limit: the server had to answer inside 24 s or Netlify cut it off (ADR-0063). Generation now runs in a background worker the app polls (ADR-0115), so a candidate past the line is slow, not unusable. Times were measured with 3 calls running at once, so they run slower than a single call on its own would.

All the numbers tokens, mean, min and max, file size

Candidate	Images	Image tokens (median)	$ per image	$ per 1,000	vs cheapest	Mean	Median	p90	Min	Max	File size
2 · lowprevious prod	28 / 28	158	$0.0353	$35	1.1×	24 s	24 s	27 s	18 s	35 s	1,505 KB
flare · lowcurrent prod	28 / 28	158	$0.0316	$32	1.0×	22 s	23 s	24 s	17 s	27 s	1,439 KB
flare · medopenai candidate	28 / 28	343	$0.0374	$37	1.2×	23 s	24 s	26 s	18 s	30 s	1,452 KB
sunburst · lowopenai candidate	28 / 28	158	$0.0316	$32	1.0×	26 s	26 s	29 s	19 s	31 s	1,420 KB
sunburst · medopenai candidate	28 / 28	343	$0.0372	$37	1.2×	27 s	28 s	31 s	21 s	32 s	1,448 KB

Time by category mean seconds per drawing type

Category	Drawings	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Freehand scenes	1	25	24	23	24	23
Coloring page, magic brush	1	30	23	22	28	28
Coloring page, colored by hand	1	22	23	25	29	29
Coloring page, barely started	1	25	23	26	27	31
Crayon captures	1	22	22	25	26	28
Filled drawings	1	22	21	20	22	24
Outlines only	1	20	18	20	20	24
Magic brush on blank paper	1	26	22	24	26	27
Messy sessions	1	19	21	20	24	23
Night mode	1	26	23	23	25	29
Pretend-play probe	1	25	24	24	30	27
Scribbled fill	1	25	22	22	26	28
A few strokes, one color	1	25	25	27	25	28
Store scenes	1	22	22	25	27	31

Calls that returned no image

none

Every drawing came back as an image from every candidate. No refusals, no errors.

Gallery

every drawing, every candidate

Each drawing is followed by one candidate per column. The compact view shows the first trial; switch to both for more detail. Tap any result to swap it for the child's drawing and back, so you can see exactly what changed. Use the toolbar to hide candidates you have ruled out.

Trials

flare

sunburst

Freehand scenes

1 drawing

Free drawings at low, medium, and high line counts.

car-hiwide

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

Coloring page, magic brush

1 drawing

A coloring page revealed with the magic brush: flat color along the strokes.

balloontall

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

Coloring page, colored by hand

1 drawing

A coloring page with regions scribbled in using palette colors.

appletall

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

Coloring page, barely started

1 drawing

A coloring page just opened, or with a stroke or two on it.

astronauttall

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

Crayon captures

1 drawing

Store scenes replayed in the live app with the crayon, captured off the real canvas.

balloontall

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

Filled drawings

1 drawing

Model-authored art with solid shapes and scribbled fill.

apple-treesquare

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

Outlines only

1 drawing

Open outlines with nothing filled in. With no color to anchor the palette, this is where models invent the most. Judge these rows for invention, not beauty.

balloontall

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

Magic brush on blank paper

1 drawing

Rainbow color revealed along strokes on an otherwise empty page.

densesquare

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

Messy sessions

1 drawing

Dots, tangles, pretend writing, and crammed corners.

corner-crammedwide

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

Night mode

1 drawing

Chalk line art on dark paper.

circletall

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

Pretend-play probe

1 drawing

A toy sword. The right answer is a picture, not a refusal.

toyswordtall

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

Scribbled fill

1 drawing

Model-authored art with areas colored in by visible back-and-forth passes.

half-colouredtall

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

A few strokes, one color

1 drawing

A handful of strokes in a single palette color, placed the way a toddler places them.

blacktall

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

Store scenes

1 drawing

The authored store-screenshot scenes rasterized onto paper: full multi-subject art.

balloontall

Child's drawing	Trial	2 · low	flare · low	flare · med	sunburst · low	sunburst · med
Child's drawing	#1
	#2

How this was measured

Inputs are built to look like what /api/generate-image really receives: the paper color, the app's coloring line art, and the child's marks in the app's 15-color palette, flattened into one image. Regenerate them with npm run model-eval:fixtures.
Every call sends the production request. The prompt and the safety instruction are checked byte for byte against the app source before the run starts. Temperature is the default.
The OpenAI candidates go through the Responses API image tool, not /v1/images/edits. Only that path accepts a real system instruction and lets the model decline in words, which is what the app turns into its safety refusal. The image tool is left optional so the model can still decline.
Cost is measured token usage times list price. Image output dominates: gemini-2.5-flash-image $30, gemini-3.1-flash-image $60, gpt-image-2 $30, gpt-image-2.5-flare $30, gpt-image-2.5-sunburst $30, gpt-image-1.5 $32, gpt-image-1-mini $8 per million image-output tokens. The OpenAI rows also include the tokens of the text model that reads the drawing and calls the image tool.
Time is wall-clock per call, with up to 3 calls in flight at once; times include concurrent load.
Safety is only spot-checked here, with one pretend-play drawing that must be allowed. The blocked-content corpus needs REDTEAM_FIXTURE_KEY and npm run redteam.
Run 2026-09-10T21-10-52-556Z-image25-winning-prompt.