$ cat wiki/entities/sakana-ai.md
Sakana AI
Latest
2026-09-11: Released Fugu Max and Fugu Ultra v2, two
2026-09-11: **Fugu Max widens the orchestrated pool to open-weight and
2026-06-22: Fugu v1 released (1 pass). Recorded as a date only
Overview
A Tokyo-based AI R&D company, founded 2023 by former Google researchers, which describes itself as a "Frontier AI R&D company" and whose stated focus is affordable generative models that work well with small datasets and are optimised for Japanese language and culture (source) (1 pass).
It is the first entity on this wiki that sells a model it did not train. Its product line, Fugu, is a learned orchestrator: a router that hands each task to a pool of other labs' models and returns one answer through one OpenAI-compatible endpoint. Sakana's own phrase is "a Multi-Agent System, Delivered as One Model" (source) (2 passes). Everything Model Routing holds treats routing as infrastructure — a library, a gateway, a classifier, an acquisition. Sakana prices it per token, as a model, and publishes benchmark scores for it against models. That is a different commercial object, and it is why this page exists.
This page exists from 2026-09-12, and its absence until now was a miss rather
than a judgement. Nothing in this repository — no snapshot, no page, no log line
— mentioned Sakana or Fugu before today, although Fugu v1 shipped 2026-06-22
(1 pass) and hardmaru, David Ha's handle, has been listed in this repo's own
sources.yaml X-account block since the file was written
(source). The daily X
sweep has been recorded as "the weaker check" in the log for six consecutive runs;
this is what that weakness cost — a tracked person's company, and a product
category this wiki has a whole concept page for, missed for eighty-two days.
Models & Products
- Fugu Max — released 2026-09-11, the cost-performance tier, $2/M input · $6/M output
- Fugu Ultra v2 — released 2026-09-11, the capability tier, $5/M input · $30/M output
- Fugu and Fugu Ultra (v1) — shipped 2026-06-22 (1 pass). No page: nothing beyond the release date was read on this run, and this wiki does not write a page it cannot cite. Named here so the version history is not silently missing.
No open weights are named for any Fugu model in anything read. The product is the endpoint (source).
Recent Activity
- 2026-09-11: Released Fugu Max and Fugu Ultra v2, two tiers of one orchestration architecture — Max optimised for cost-performance, Ultra v2 for capability on complex multi-step tasks. Fugu Max is claimed best overall on six benchmarks (Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, SWEFish) and to expand the cost-performance Pareto frontier on 7 of 10 benchmarks, at output pricing 40–60% lower than Sonnet 5, GPT 5.6 Terra and Kimi K3. Fugu Ultra v2 is claimed best or joint-best on five of eight benchmarks, with Chartography 48.3 against Claude Opus 5 27.3, and DeepSWE 74.3. Why it matters: the v2 claim is not the score but the pool — Sakana states the scores are reached "without Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool", i.e. an orchestrator built from weaker parts claiming to beat the parts it removed. What is not established: no independent reproduction, no enumeration of the pool, no latency figure, no license, and no per-request cost decomposition (source)
- 2026-09-11: Fugu Max widens the orchestrated pool to open-weight and specialised models, "including NVIDIA Nemotron family through our collaboration with NVIDIA" — see NVIDIA and Nemotron 3.5 Lightning. Sakana's stated thesis: "Open models become dramatically more useful when orchestrated together rather than used in isolation" (source) (2 passes)
- 2026-06-22: Fugu v1 released (1 pass). Recorded as a date only — nothing else about it was read on this run (source)
- 2026-03: Investment from Mitsubishi Electric Corporation (source) (1 pass)
- 2025-11-17: Series B of ¥20 billion (≈ $135 million) at a $2.65 billion post-money valuation, up from $2.5 billion pre-money. Total raised $379 million across 7 rounds; 250 employees as of July 2026 (source) (1 pass)
Strategic Position
Sakana competes on assignment, not on capability. Every other lab on this wiki sells the output of a training run; Sakana sells the decision about whose training run answers your request. The two positions have opposite exposure to the same event: a cheaper, better frontier model is a threat to a lab and an input to Sakana.
That is also the weakness. The Fugu tiers' published benchmark scores are
system scores, and Model Routing already records why that is not
the same claim as a model score — the number characterises an orchestrator over a
pool, and the pool is not enumerated in anything read. A model page's Pricing
row describes what the caller is charged; for Fugu it does not describe what the
request costs to serve, and nothing read states who absorbs the difference.
The v2 pool claim is the position stated as a product decision. Removing Claude Fable 5, Claude Fable 5.1 and GPT-6 Astra from the pool and still claiming frontier output is an argument that the orchestrator, not the frontier model, is where the capability lives. It is an argument this wiki cannot check: no independent reproduction exists, and there is no Artificial Analysis page for any Fugu model (1 pass) — so the one third-party instrument this wiki routinely quotes has nothing to say about it.