Sakana Fugu ULTRA Review (Better than Fable 5?!)

Sakana Fugu Ultra is an orchestrator, not a frontier model — and it eats your quota on a single prompt.

## What Fugu Ultra actually is Sakana Fugu is a "learned multi-agent orchestrator that routes across a pool of strong models like GPT, Claude, and other frontier models." The default fugu tier is tuned for everyday coding and chat; fugu-ultra uses a deeper pool and "one to three agents per task to maximize quality on really complex tasks at the cost of higher latency and much higher per token price." On Sakana's own (self-reported) benchmarks, Ultra "broadly matches or slightly beats current frontier models including Mythos preview as well as Fable 5" on hard coding and reasoning — but the results "are still mostly self-reported" as of launch day.

## The cost problem is real The creator is on the $20/month standard plan. One prompt — a simple project directory review — used "29%" of the weekly allowance, then a second Codex run chewed through "84% for the 5-hour usage limits." Pro is $100/mo, but "is it really worth it when we already have access to so many good strong models right now and we can just manually route different models?" The honest framing: "it's really about convenience."

## Setting it up in Codex Drop your API key into ~/.bashrc, source it, then point a custom provider at base URL api.sakana.ai/v1 and set the default model to fugu-ultra. You'll get two effort levels: high and xhigh. The creator used xhigh on a trivial task and called that "a mistake on my part."

## Use the right effort, or get rekt xhigh = "maximum reasoning effort," worth it only for "code refactors or something really, really difficult, a long horizon task." For startup work — reviews, system checks, drafting features — use high. The Fable 5-style output you see is just "a compilation of GPT and Claude models taking a look at the task and then coming to a conclusion," so "it comes down to how well you structure your project and how good your prompt is."

keyPoints