{"rewrite":{"id":"r_9b8c8ec342eba1c8b04f5fac","clusterId":"c_81845026987154919ed6ce72","slug":"sakana-ai-releases-fugu-ultra-v2-and-fugu-max","model":"deepseek-v4-flash","headline":"Sakana AI Releases Fugu Ultra v2 and Fugu Max","summary":"Sakana AI, based in Tokyo, announced Fugu Ultra v2 and Fugu Max, systems that allocate tasks across multiple AI models. Fugu Ultra v2 outperformed the model 5.1 and GPT-6 Astra on multiple benchmark tests, though neither is among the models it uses. Fugu Max offers a wider variety of usable models and kept costs down while scoring highest on Terminal Bench 2.1 compared with the other models shown.","whyItMatters":"Sakana AI's routing approach is being priced and benchmarked directly against single-model rivals, which is where multi-model orchestration either holds up commercially or does not.","webCardHtml":"\u003cp\u003eAccess runs through pay-as-you-go API billing or a subscription. Per million tokens, Fugu Ultra v2 costs $5 for input, $0.50 for cached input, and $30 for output; Fugu Max costs $2, $0.25, and $6. Monthly subscription tiers are $20 for Standard, $100 for Pro, and $200 for Max.\u003c/p\u003e\u003cp\u003eSakana AI positions Fugu Max as the more economical option in the series, with more models available to it than its siblings.\u003c/p\u003e","blueskyPost":"Fugu Ultra v2 beat GPT-6 Astra on benchmarks without using it or the model 5.1. Sakana AI's routing layer is doing the work, not any single model inside it.","twitterPost":"Fugu Ultra v2 beat GPT-6 Astra without running it. Sakana AI's edge is the routing layer, not any one model.","threadsPost":"Fugu Ultra v2 outperformed GPT-6 Astra and the model 5.1 on benchmarks, and it runs on neither. Sakana AI's Fugu Max instead scores highest on Terminal Bench 2.1 while keeping costs down. The variable being tested is model orchestration, not raw model strength.","newsletterBlurb":"Sakana AI announced Fugu Ultra v2 and Fugu Max, systems that split tasks across multiple models. Ultra v2 surpassed the model 5.1 and GPT-6 Astra on several benchmarks, even though it does not use those models. Fugu Max offers more usable models and posted the highest Terminal Bench 2.1 score among the compared models at a lower cost.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20260914-fugu-ultra-v2/\",\"title\":\"Sakana AI Releases 'Fugu Ultra v2,' Which Surpasses GPT-6 Astra by Switching Between Multiple Models, Plus the Cost-Efficient 'Fugu Max'\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":4806,"outputTokens":611,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1789361975,"createdAt":"2026-09-14T04:55:45.000Z","publishedAt":"2026-09-14T04:58:24.000Z","updatedAt":"2026-09-14T04:58:24.000Z"},"cluster":{"id":"c_81845026987154919ed6ce72","canonicalTitle":"Sakana AIが複数モデルの使い分けでGPT-6 Astra超えを達成する「Fugu Ultra v2」をリリース＆コスパに優れた「Fugu Max」も登場","representativeArticleId":"a_ef039a432aea048a7eda9c35","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[],\"studios\":[],\"people\":[],\"type\":\"release\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-09-14T03:52:00.000Z","lastSeenAt":"2026-09-14T03:52:00.000Z","updatedAt":"2026-09-14T04:58:20.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20260914-fugu-ultra-v2/","title":"Sakana AIが複数モデルの使い分けでGPT-6 Astra超えを達成する「Fugu Ultra v2」をリリース＆コスパに優れた「Fugu Max」も登場"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":[],"studios":[],"people":[],"type":"release","domain":"other","is_roundup":false},"keyFacts":null}