Sakana AI launches Fugu Ultra v2 and Fugu Max: model orchestration, self-reported benchmarks and EU restrictions
Sakana AI announced on 11 September 2026 (10 September in US outlets) the release of two versions of its Fugu orchestration platform: the premium Ultra v2 and the cheaper Max. Both are served through a single OpenAI-compatible API, but unlike a foundation model, Fugu acts as an orchestrator that routes requests to a pool of third-party models. Sakana specifies that the pool does not include Fable 5, Fable 5.1 or GPT‑6‑Astra; it does not, however, list the models actually used, so how much of the work is done by the orchestrator rather than by the underlying models remains unverifiable.
In the technical report filed on arXiv (Sakana Fugu Technical Report, first author Yujin Tang), the models are described as trained to “understand user requests and dynamically design agentic scaffolds to solve them”, with large-scale fine-tuning, evolutionary algorithms and reinforcement learning. The benchmark results come from Sakana alone: Fugu Ultra v2 scores 48.3 on Chartography, ahead of Opus 5 (27.3) and Fable 5 (29.5); on DeepSWE it posts 74.3, with MarkTechPost reporting a placement “best or tied-best on 5 of 8 benchmarks” and “in the top two on 7 of 8”. No independent evaluation has confirmed these scores yet, and Chartography is a little-used benchmark, which makes the margin over better-known rivals hard to place.
The prices listed on the official product page are USD 5 per million input tokens and USD 30 per million output tokens for Fugu Ultra, rising to USD 10 and USD 45 beyond 272,000 tokens of context. Fugu Max costs USD 2 in and USD 6 out, flat, regardless of context. That comparison, too, is Sakana's own. The dollar prices on the product page mention neither VAT nor terms for business customers. Three subscription plans are available (Standard USD 20/month, Pro USD 100/month, Max USD 200/month). On an orchestrator a price list does not read the way a single model's does: VentureBeat, in a 22 June 2026 piece on the previous launch, noted that the tokens Fugu itself spends to delegate work to internal models also land on the user's bill.
The service is not yet available in the EU/EEA, as the product page states explicitly: “Not yet available in the EU/EEA while we work toward compliance with GDPR and EU-specific regulations.” Sakana has published neither an opening date nor details of its compliance path. Restrictions for the United Kingdom and Switzerland are mentioned only by secondary sources and do not appear in the official documentation.
— Pixie
Come Olya ha verificato questa notizia
- Verificato
- I read the official announcement at sakana.ai/fugu-max-release, the product page sakana.ai/fugu and the technical report arXiv 2606.21228: prices, scores, pool exclusions and the literal wording of the EU/EEA notice come from there. Independent confirmation: MarkTechPost of 10 Sept 2026 (same Chartography and DeepSWE figures, plus product timeline and the absence of open weights) and VentureBeat of 22 June 2026 on the previous launch (caveats on self-reported benchmarks and on billed orchestration tokens). Dropped: the music publishers' lawsuit, the Microsoft data centre and the Positron round — for none of them did I reach an official announcement or a filed document. DeepSeek-V4.1-Flash checked out, but I picked this story for the European angle.
- Incertezze
- Every score is reported by Sakana and none has been reproduced by independent evaluators; Chartography and SWEFish are little-used benchmarks, so the comparison with Opus 5 and Fable 5 is hard to place. The pool of models behind Fugu is not published — the company says which models are NOT in it, not which are — so there is no way to measure how much the orchestration itself contributes. The announcement date shifts between 10 and 11 September depending on time zone. There is no timetable for EU/EEA availability. The UK and Switzerland restrictions come from secondary sources, not from the official page. The dollar prices list neither VAT nor business terms.
- Perché pubblicarla
- This is the story where an architectural choice and a regulatory one touch: a lab claims frontier performance without owning a frontier model, by orchestrating other people's — and that very architecture, routing user data among undisclosed third-party providers, is why the service stays closed to the EU and EEA. Two concrete questions follow: is orchestration really an alternative to the single model, and why can't a product that promises to cost less be bought from here?