Wait Which Model?
← Back to directory
Sakana AI

Fugu Max

Released Sep 11, 2026 · knowledge cutoff 2026-08

Status
Unknown
Location
Japan
Modality
Multimodal
Context window
1M
Max output
128K
Speed
Price ($/MTok in / out)
$2 / $6
Cost per task
Open weights
No
Benchmarks0 of 10 reported

MMLU-Proretired

GPQA Diamond

SWE-bench Verified

SWE-bench Pro

Terminal-Bench 2.1

AIME (latest)retired

Humanity's Last Exam

LMArena Elo

GDPval-AA v2

ARC-AGI-2

Strengths & weaknesses

Strengths

  • Genuinely cost-competitive on output pricing against same-class flagship single models rather than just against other orchestrators
  • One OpenAI-compatible endpoint handles function calling, structured outputs and built-in web search/fetch without extra integration work
  • Sakana reports it as best-in-class on several of its own coding and document-reasoning evals against single frontier models, not just against its own prior tier

Weaknesses

  • Routing is a black box: which underlying model answered a given request is not exposed or auditable, a real gap for compliance-sensitive use
  • Orchestration overhead shows up as billed tokens the user never sees — one hands-on test logged the large majority of a coding task's billed tokens as back-channel coordination rather than visible output
  • Latency is unpredictable rather than merely high: simple questions have taken well over a minute end to end while the coordinator deliberates regardless of task difficulty
  • Too new for independent benchmark verification — every headline number so far is Sakana's own vendor-reported evaluation

Not a single foundation model: a language-model orchestrator (built on Sakana's TRINITY and Conductor research, ICLR 2026) trained to route each request across a fixed, undisclosed pool of open-weight and specialized models — including NVIDIA's Nemotron family — and to recursively call itself, then assemble a final answer behind one API. Closed/proprietary: no weights are published and the composition of the model pool is not disclosed. Cache-hit input is $0.25/M; web_search and web_fetch tool calls are billed separately, reported as $0.007/call by some trackers and as $10/1,000 calls ($0.01/call) by OpenRouter's listing — this entry could not reconcile the two. Sakana's launch post claims best-overall score on six of its own benchmarks (Terminal-Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench and SWEFish) against single frontier models, but the actual percentages are rendered only in chart images in the announcement, not as text, so none of this site's tracked benchmark fields could be filled with a specific number; GDP.pdf (a Surge AI PDF/document-reasoning benchmark) is a different, unrelated benchmark from this site's gdpvalAA (Artificial Analysis' GDPval-AA v2), so no mapping was made between them. No Artificial Analysis Intelligence Index coverage exists for any Fugu model as of launch, hence costPerTask and speed are null. Knowledge cutoff (2026-08) is sourced from OpenRouter's model listing, not confirmed on an official Sakana page — treat as third-party-sourced.

For developers

API model strings

  • sakanafugu-max
  • openroutersakana/fugu-max

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor

News