Sakana Fugu Ultra Tests (high vs xhigh reasoning)

Fugu Ultra high reasoning completes four tasks on the 5-hour Codex window, but three of four outputs underwhelm.

## Cost vs. xhigh The previous xhigh run burned 84% of the 5-hour Codex usage limit on a single prompt. Switching to high reasoning effort let the creator run four full projects in roughly the same window and only then approached 100% of the limit. Token spend is still high, but high is the better lever for Codex users.

## The one win: multi-agent visualizer The standout output was a single-file HTML visualization of Sakana's own orchestration: planner → research → builder → sandbox → critique (adversarial loop) → supervisor, with a live routing log in monospace, animated dark canvas, and latency counters. The prompt specified a dark theme canvas, live log console, and latency counters, and the model delivered without revision.

## The three misses 1. Cyberpunk 3D Hong Kong — built on a 4 GB VPS with no GPU; the model could not realize the neon/easter-egg game properly. Hidden click-the-moon singularity overlay exists but is nearly impossible to find. 2. Astro.build clone — UI/UX was close, but color/theme was weak. The creator explicitly says GLM 5.2 produced a better design even with just a text-model pipeline (Kimi 2.7 → GLM 5.2). 3. Original site from spec — Tailwind + Google Fonts + offline requirement conflicted with the CDN image load, so the hero fell back to a default template.

## Practical guidance Fugu Ultra on Codex needs PRDs, context files, and structured prompts — one-shot prompts fail unless they are extremely detailed. To cut cost further, a community tip confirms Fugu Ultra works inside Pi and Droid, dropping token spend below Codex. Next test will be the cheaper Fugu default, which only matches Ultra on SWE-bench Pro and trails on the rest.

## Recommendations - Stick to high over xhigh on Codex for Fugu Ultra. - Run 3D/WebGL projects locally on Windows with a GPU — skip VPS. - Skip Fugu Ultra when an existing GLM 5.2 pipeline already handles design work. - For Fugu, supply a PRD, context files, and clear offline constraints up front. - Try Fugu Ultra through Pi or Droid for lower token cost than Codex. - Wait for the Fugu default test before deciding if cheaper reasoning is worth the coding trade-off.