Gutter is the cartoonist behind Igor's blog raccoons. It draws in one specific look — a plush, soft-vinyl "designer toy" render — and its two recurring characters have a strict canon. This page hands the same fixed scene to nine image models and asks a simple question: how close does each get to how Gutter really draws?
The scene is Larry and Igor together in the Den. Gemini 3 Pro, run through Gutter's own pipeline, is the house baseline — the others are measured against it. Every render carries its real per-image cost (billed usage.cost, not the catalog rate), its latency, and a five-point canon check.
Built with the five-block ordering from Gutter's recipe: layout first, then the render style flagged as most important, then Larry (beard + one claw called out), then Igor (shirt text quoted), then the set. The four Gutter character/style reference images ride along on every call.
| Model | Provider | $ / image | Latency | Canon check |
|---|---|---|---|---|
| Gemini 3 Pro baseline | Google direct | ~$0.134 list* | 42.2 s | ✓✓✓✓✓ |
| GPT-Image 2.5 flare | OpenRouter · openai | $0.255 | 61.3 s | ✓✓✓✓✓ |
| GPT-Image 2.5 sunburst | OpenRouter · openai | $0.097 | 47.4 s | ✓✓✓✓✓ |
| Gemini 3.1 Flash | OpenRouter · google | $0.070 | 10.7 s | ✓✓✓✓✓ |
| Seedream 5.0 Pro | OpenRouter · bytedance | $0.099 | 119.5 s | ✓✓✓✓✓ |
| Qwen-Image 3 | OpenRouter · qwen | $0.036 | 59.6 s | ✓✓✓✓✓ |
| Grok Imagine 2.0 | OpenRouter · x-ai | $0.080 | 19.5 s | ✓✓✓✓✓ |
| Recraft v4 styles | OpenRouter · recraft | $0.040 | 29.9 s | ◑✓✓✓✓ |
| Muse Image | OpenRouter · meta | $0.010 | 26.1 s | ◑✓✓✓✓ |
* Google's direct image API returns no billed-cost field; ~$0.134 is the published Nano-Banana-Pro per-image list rate (this render used 1,120 image output tokens). Every OpenRouter figure is the real billed usage.cost.
This is the bar. Warm-brown chibi plush, brown lace-ups on Larry, rainbow glasses on Igor, shirt text crisp, blue+yellow Crocs. Gutter's actual generator; everything else is compared to it.
Best-looking of the challengers. The richest felt texture and the only render besides the baseline to land the claw on Larry's actual left arm. Also the priciest and slowest.
Flare quality at a third of the price. Cozy, tactile, canon clean across the board — shirt text and Crocs both verified on a zoom. Strong value pick.
Fastest full pass — 10 seconds. A hair flatter than the Pro baseline but unmistakably the same house look, with every canon detail intact. The everyday workhorse.
Clean render, slow to arrive. Held every canon point and the plush look, but took two minutes — the slowest of the lot by a wide margin.
Cheap and on-model. Caps at three reference images, so it ran on two — and still nailed the canon, even adding a raccoon-anatomy sketch on the wall. Nice attic mood.
Fast and faithful. Rejected the aspect-ratio parameter with four refs, passed cleanly with two. Full canon, warm plush look, under 20 seconds.
Canon held, style drifted. Every character detail is right, but it painted a flatter, more illustrated look and boxed the scene in a black comic-panel border the prompt never asked for.
Cheapest, and the best raw felt. Stunning needle-felted texture — but it re-read the raccoons as naturalistic grey animals instead of Gutter's warm-brown chibi palette, and dropped Igor's oversized head. Off the house model, one cent.
The reference images do the heavy lifting. All nine models held the four character canon points — exactly one red claw, the grey-white beard, "TECHNOLOGIST" spelled correctly, and mismatched blue+yellow Crocs. With four strong reference images attached, even the cheapest model kept the characters on-model. That is the real headline: canon is a reference-image problem, not a prompt-wording problem.
Where they split is the look. Seven of nine reproduced Gutter's warm-brown chibi plush style. Two held the canon but drifted: Recraft flattened the render and added an unrequested comic border, and Muse — while producing the most convincing felt texture of all — re-drew the pair as naturalistic grey raccoons rather than the house palette.
Chirality is still unsolved. The claw belongs on Larry's left arm. Only the two GPT-Image renders placed it there; every other model, including the Gemini Pro baseline, mirrored it onto the wrong arm. Passing a reference that shows the correct side does not fix it — a known limitation of these models.
One prompt, one scene, nine models. The baseline runs through Gutter's real pipeline: gemini-3-pro-image-preview via the gen-image skill on a Google API key, the same path every blog raccoon is drawn with. The other eight run through OpenRouter's image endpoint (POST /api/v1/images).
Total spend to build this page: about $0.96 across all ten renders (nine models plus one baseline re-run for timing).