Sakana says Fugu matches Fable 5. I tested it. | build in public

On June 22 a lab in Tokyo released a 7 billion parameter model that claims to stand shoulder to shoulder with Fable 5, the most powerful AI in the world, and the one the US government switched off everywhere back in June. A fraction of the size, a lower price, and no export controls. Those are big claims, so I tested them. The trick is that Sakana Fugu isn't a frontier model at all. It's an orchestration model. It sits on top of the models you can still reach, Opus 4.8, GPT-5.5, Gemini, hands each part of your task to whichever one it thinks is best, makes them check each other, and feeds you back one answer. One model quietly commanding all the others. Fable 5 is the child prodigy who just knows the answer. Fugu is the group of clever kids working together. They are not the same thing, but Sakana is comparing them directly, so I wanted to see if that holds up. So I did the thing the benchmarks can't do for you. I put Fugu Ultra head to head with Opus 4.8 on two real tasks, in the same harness, and paid for both out of my own pocket. A simple one-line bug fix, and a harder design task I could judge with my own eyes. The results flip depending on the job, and that's the whole video. On the bug fix both models landed the identical fix, except Fugu cost almost six times more to get there, because it hired a team for a job one model does on its own. On the design task it flipped the other way, Fugu came in cheaper and faster, and you could see where the saved money went. There is even a button in its result that tells you to drag to steer the heat, and when you drag, nothing happens. An open-source model I covered last week arguably did the whole thing better for less. Where I land is that there's no clear winner, and I think that's the more interesting answer. But zoom out and the bet underneath this is real. For years the race was about who could train the biggest model. Orchestration is a bet that the edge moves to whoever coordinates the models we already have, and that matters a lot more in a world where the best models keep getting switched off. The idea is the story here, even if this first version is oversold. For most things you're fine sticking to a single model. It's cheaper and saner. But this stuff changes in a week or two, so don't take my word for it, put a few dollars down and test it on your own work like I did. CHAPTERS A lab in Tokyo just shook up the AI race One tiny model to command them all What an orchestration model actually is The claim: Fable 5 performance for less Where Fugu really lands on the benchmarks Why I'm not writing it off The catch: you only ever see the bill So I tested it against Opus 4.8 The setup (Claude Code and the AI gateway) Test 1: a one-line bug fix Test 2: KebaBench, my favourite test Opus 4.8's kebab Fugu Ultra's kebab (and a button that does nothing) So who actually won? Why orchestration might be the future What I'd actually do (and test for yourself) SOURCES & FURTHER READING Sakana's official Fugu release (the pitch and the Fable comparison): https://sakana.ai/fugu-release/ Sakana Fugu product page (pricing, variants, the architecture diagram): https://sakana.ai/fugu/ The Fugu technical report (how the orchestration actually works, TRINITY and the Conductor): https://arxiv.org/html/2606.21228v1 VentureBeat's benchmark breakdown (the SWE-Bench Pro numbers in the video): https://venturebeat.com/orchestration/no-claude-fable-5-no-problem-sakana-achieves-frontier-performance-with-new-fugu-multi-model-auto-synthesis-system What Fugu Ultra is, explained: https://wavespeed.ai/blog/posts/what-is-fugu-ultra/ The GLM-5.2 vs Qwen thread I referenced: https://www.reddit.com/r/LocalLLaMA/comments/1ua1na0/whats_more_impressive_glm_51_52_or_qwen_35_36/ My GLM-5.2 video (the open-source model that beat Fugu on the kebab test): https://www.youtube.com/watch?v=CEPBYfRcoBU SOCIALS Follow me on X: https://x.com/robj3d3 Follow me on Insta: https://instagram.com/robhallam_ Subscribe to my free newsletter (weekly tips from founders using AI): https://getmyfirstdollar.com