Home / Vaults / Infographic bake-off
Infographic bake-off: every image model on OpenRouter, judged blind, 10 October 2026
Which image model should the newsroom use for an infographic, for which job, and at what cost? Every image-output model on OpenRouter on the day, its auto-router and our own code-drawn renderer as a control were given the same briefs from one article: an 18-word stat card, a concept slide, a diagram with exact labels, a chart with fourteen numbers, a 240-word document, the whole 2,800-word article, an edit of their own image, a UI component, a brand slide, a three-slide deck, and finally the conclusion of the bake-off itself. Five rounds, models cut after each. 101 images, every one judged blind by a Designer agent, every cent recorded by OpenRouter: $10.44. It was published with the article The infographic bake-off.
Read key:
sgit_public_read_89ef38c8db580a5c858ae1889859f30792b2a16fdf370d58a8cd8e58000cc6be:po5i477iIn the official UI: open it read-only in a new tab · From the CLI:
sgit clone sgit_public_read_89ef38c8db580a5c858ae1889859f30792b2a16fdf370d58a8cd8e58000cc6be:po5i477iPublished deliberately under the
sgit_public_read_ prefix, and derived one-way from a vault key kept in the gitignored tier and never published. Classified with check_credential.py before it touched this page, cloned back with the published key alone and compared with the source file by file, and checked with an all-zeros read key as the negative control. Created, pushed and audited with sgit-ai 0.20.0. The OpenRouter key the run used was read from a file outside the vault and appears nowhere in it.See it live, here
The vault opens as an app: the overview, each round with every image, its cost, time, fact check and the judge's reason, a page per model, the guidance, the ledger of every generation, and the method and roles. You can also open it in the official UI.
The one idea
Choose the model by the job, not by the budget. A cheap model at four cents an image makes concept slides, charts and consistent decks as well as models four times its price, and fails every brief where each character is specified. The most expensive model is the only one that turns a whole document into a correct infographic and the only one that edits an image without breaking something else. The middle one is the fast all-rounder.
What is in it
Every image, side by side, with its cost and the judge's reason
For each brief, every model's image with its recorded cost, seconds and size; the Designer's fact check of the required text, OCR recall as an independent check, anything invented; the overall score; and one line on why. Click any image for full size.
Edits, components and brand rules
Round 3 asked each model to change three things in its own diagram and nothing else, to draw a UI component with seven exact rows, and to follow house brand rules. Only GPT-5.4 Image 2 made the edit without breaking anything; the cheap model had no usable image in the round.
Which model for which job, and how to write the brief
For ten kinds of job: the model to use, the fallback, what to avoid, the cost and time, and the evidence. Twelve prompt rules learned from the faults (never name the platform, say "No other text", forbid invented numbers, attach slide 1 to every later slide of a deck), and how to call the models through OpenRouter.
Every generation, every cent
One row per image: round, brief, model, OpenRouter's recorded cost, seconds, size, fact check, OCR recall, overall and usable scores. The Accountant's notes add cost per usable image, what drives cost, the cost of an edit and monthly projections.
How it is built
The briefs are in scenarios.json, built from the Re-anchoring article and its Five readers views (source/). tools/run.py calls OpenRouter's chat completions with image output, in parallel, and records each generation's cost, tokens, seconds, size and id; tools/ocr_score.py reads every image with tesseract in four segmentation modes and checks the required strings; Designer agents judged copies with random ids, never seeing the model; tools/ledger.py unblinds and totals; the Accountant and the Data Scientist, each an agent with a role file in roles/, wrote analysis/. The images are kept as full-size WebP in img/; the original PNGs are not in the vault.
The audit, honestly
What was scanned. Every file before the first commit, for credential shapes (vault keys, every sgit_ credential prefix, private-key headers, the SG/Send access token) and for the OpenRouter key and its prefix; and again after cloning with the published read key alone. The clone matched the source, file by file.
What was found. Nothing. The negative control, an all-zeros read key against the same vault id, returned nothing.
What it does not claim. Most cells hold one or two images, and repeats of one brief moved scores by up to two points, so a one-point difference between two models is noise; the tiers are not. The judge is one kind of agent with one rubric. Prices and models are OpenRouter's on 10 October 2026 and will change. Nothing here is a statement about any vendor beyond what these images show.
Write-key status: escrowed, in the gitignored credential tier, before this page was written.
Derived facts
From admin/build/catalogue_derive.py po5i477i <read key hex>, read-only, no token, no clone.
- Files: 378 · plaintext size: 8818 KB
- Commits: 3 · last updated: 2026-10-10 · HEAD:
obj-cas-imm-953d35c76215 - Top level:
README.md,analysis/,app.json,app/,img/,index.html,models.json,roles/,runs/,scenarios.json,scores/,source/,tools/ - Vault app: yes, entry
index.html· browser-renderable: yes
Notes
Where this came from. A voice memo by Dinis Cruz asking for a bake-off of image models, from cheap to expensive, with the cost, the time and the ability to follow the brief recorded, ending in three models and guidance; the source article Re-anchoring and its Five readers views; the Designer role of the SG/Send agent team, and an Accountant and a Data Scientist written for this run.
Who wrote this. agent@riskmandate.ai (Claude Opus 5.5, claude-opus-5-5), in the sgit.ai site session, with Claude agents for the judges and the analysts. AI-generated text, disclosed as Article 50 of the EU AI Act asks; the person with editorial responsibility is Dinis Cruz.