Sakana AI Unveils Fugu: a Multi-model Task Routing System
WHO:
MarkTechPost - Sakana AI, an AI startup.
VentureBeat - David Ha, CEO and co-founder of Sakana AI, formerly of Google Brain.
WHAT:
VentureBeat - Fugu Ultra scored 73.7 on SWE-Bench Pro, outperforming Anthropic's Claude Opus 4.8 (69.2) and OpenAI's GPT-5.5 (58.6).
VentureBeat - Fugu Ultra narrowly edged out Claude Opus 4.8 (50.0 vs 49.8) on Humanity's Last Exam but fell short of Fable 5 (53.3).
VentureBeat - On the MRCRv2 long-context-recall test, OpenAI's GPT-5.5 maintained the lead with a score of 94.8 compared to Fugu Ultra's 93.6.
MarkTechPost - Sakana AI launched Fugu, a multi-agent orchestration system.
VentureBeat - Fugu routes tasks across a swappable pool of specialized AI agents, achieving frontier-level AI performance.
VentureBeat - Fugu Ultra, a variant of Fugu, is engineered for complex tasks and matches industry-leading models on benchmarks.
VentureBeat - Fugu exceeds the performance of Anthropic's Claude Fable 5 on LiveCodeBench with scores of 93.2 for Fugu Ultra and 92.9 for Fugu, compared to 89.8 for Fable.
VentureBeat - Fugu Ultra scored 95.5 on GPQA-D, matching the performance of Claude Mythos Preview.
WHERE:
VentureBeat - Fugu is restricted from operating within the European Union (EU) and European Economic Area (EEA).
WHEN:
VentureBeat - Sakana AI's Fugu was launched in June 2026.
VentureBeat - A free second month for users who subscribe to any tier is available until July 31, 2026.
WHY:
VentureBeat - Fugu was designed to provide resilience against vendor lock-in and geopolitical export controls.
VentureBeat - David Ha stated that relying on a single company’s model for national infrastructure is a massive risk due to potential sudden access loss.
HOW:
VentureBeat - Fugu Ultra utilizes a fixed pricing structure per one million tokens: $5 for input, $30 for output, and $0.50 for cached input, with rates increasing for extreme workloads above 272K tokens.
VentureBeat - Fugu operates by dynamically routing queries to a pool of specialized AI agents, breaking down complex requests and delegating sub-tasks.
VentureBeat - Fugu is offered as a commercial, proprietary API service, with specific routing information hidden from users.
VentureBeat - Fugu charges a dynamic rate based on the specific underlying models activated, with Fugu Ultra priced at $5 per million input tokens and $30 per million output tokens.
(UPDATED June 22, 2026 21:45 UTC)
In a recent performance evaluation, Fugu Ultra achieved a score of 73.7 on the SWE-Bench Pro, surpassing Anthropic's Claude Opus 4.8, which scored 69.2, and OpenAI's GPT-5.5, which scored 58.6. While Fugu Ultra narrowly edged out Claude Opus 4.8 on Humanity's Last Exam with a score of 50.0 to 49.8, it fell short of Fable 5, which scored 53.3. Additionally, in the MRCRv2 long-context-recall test, OpenAI's GPT-5.5 led with a score of 94.8, compared to Fugu Ultra's 93.6, and Fugu Ultra offers a fixed pricing structure of $5 for input and $30 for output per one million tokens.
--
(UPDATED June 22, 2026 20:08 UTC)
Sakana AI has launched Fugu, a multi-agent orchestration system designed to route tasks across a swappable pool of specialized AI agents, achieving impressive performance levels. Fugu Ultra, a variant of the system, excels in complex tasks, outperforming Anthropic's Claude Fable 5 on LiveCodeBench with scores of 93.2 and 92.9, respectively, and matching Claude Mythos Preview with a score of 95.5 on GPQA-D. While Fugu is available as a commercial API service, it is restricted from operating within the European Union and European Economic Area, and users can take advantage of a free second month for subscriptions until July 31, 2026.