Every comparison in the 2026 Recreation rests on a choice that gets buried fast: which models do we actually run? Mad Bitcoins' archive is 9,481 generations spread across roughly 1,662 distinct prompts. Recreating even a slice of that work means committing to a model lineup, paying for it in attention and account quota, and accepting that whatever differences appear in the side-by-side grids are differences in the tools we picked, not the only tools that existed in May 2026.
So on 2026-05-18, before any of the pilot prompts ran, we built a one-prompt recon page. The test was "cartoon raccoon samurai with katana, in the style of detailed digital illustration" — adjacent to but not identical to the most-iterated prompt in the whole archive ("cartoon racoon samurai" at 64 reruns; see the prompt history report). Six free-tier image tools were aimed at that single prompt. The page lined the outputs up next to Thomas's 2022 MidJourney v4 reference, and the lineup decision was made off of one glance.
Who showed up, who didn't
The six tools tried, in the order they were sent the prompt:
The pilot lineup that came out of this is the five-column grid that powers every prompt page in the 2026 site: MidJourney (archive) plus Gemini, AI Studio, HF FLUX, and ChatGPT. The sixth column — local mflux — is the one absence worth naming in print, because it would have been the only model that wasn't a hosted service with a quota and a content policy. The disk-full failure was undramatic and entirely material: 460 GB host, 93% used, and the smallest non-gated FLUX checkpoint that mflux would actually run was bigger than the free remainder.
What the recon was actually for
The recon page didn't pretend to be art criticism. It existed to answer four boring procurement questions before the pilot committed any time to recreating eleven collections:
1. Which free tiers will sit still for a Chrome-MCP loop? Gemini and AI Studio both did. ChatGPT needed a paid Plus account and a browser-automation rig but tolerated it. FLUX was friendly to anonymous traffic but queue-dependent.
2. What's the per-image latency? 25 to 70 seconds across the surviving four. Multiplied by 109 prompts × 4 models, that's the budget the recon was sizing.
3. How distinctive are the model fingerprints? Distinctive enough. Gemini went cinematic, AI Studio went portrait, FLUX went sketch-adjacent, ChatGPT went painterly. The grid had something to show.
4. Will any of them flat-out refuse a Mad Bitcoins prompt? The raccoon-samurai sailed through all four. The real refusals — Satan, the Capitol riot, Howard Stern — surfaced later, mid-pilot, and got their own writeup in trademark theater and the refusals page.
None of those questions need to be answered in an essay. They needed to be answered before the comparison grid existed. The recon page is the seam between "we should try this" and "we shipped 109 prompts." It is also the only place in the project where a model that didn't make the cut is still visible — a row in the scorecard with action needed on it. The 2026 Recreation is the rig minus its quiet rejection.