Fugu Ultra v2.0 (Sakana AI)

Fugu Ultra v2.0, created by Sakana AI, is positioned as a multi-agent orchestration engine rather than a single monolithic model. Its architecture is grounded in learned model coordination, drawing on two ICLR 2026 papers: TRINITY, which uses an evolved LLM coordinator, and Conductor, which applies reinforcement learning to discover natural-language coordination strategies. By routing tasks through a swappable pool of open-weight and specialized models, Fugu Ultra v2.0 is designed to avoid dependency on any single proprietary frontier model, allowing it to adapt the composition of its underlying components per query.

In practical terms, this orchestration approach means Fugu Ultra v2.0 is intended for workflows that benefit from dynamic task routing and ensemble-style reasoning across heterogeneous models, rather than a single end-to-end neural network. The design aligns with a broader strategy of decoupling performance from proprietary lock-in, making the system well suited to enterprise and developer use cases that require flexible, API-compatible coordination of multiple model capabilities within a single product surface.

LLM Gatewaysakana/fugu-ultra-v2.0fugu

Open provider page Provider docs

Quick Info

Provider: LLM Gateway
Model key: sakana/fugu-ultra-v2.0
Release date: Sep 11, 2026
Last updated: Sep 11, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $5.00
Output token cost: $30.00

Limits

Output tokens: 1,000,000 tokens
Context window: 1,000,000 tokens

Transparent token rates

Compare Fugu Ultra v2.0 (Sakana AI) pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Fugu Ultra v2.0 (Sakana AI)

LLM Gateway

Coverage

A September 18, 2026 weekly AI roundup by Elena Daehnhardt groups Sakana AI's Fugu orchestrators among three contrasting model-system releases that landed on September 11, alongside Shanghai AI Lab's open-weight Atria Dawn Preview (744B, MIT licence) and Moonshot AI's hosted Kimi K2.8 Preview. The piece frames Fugu as The same roundup situates Fugu within a broader news cycle that also covers Gemini 3.8 Live Extended Thinking, SemiAnalysis measurements of NVIDIA's Vera Rubin NVL72 rack, Dario Amodei's "We Must Pace the Frontier" proposal, and CAD 150 million / EUR 100 million commitments from Canada and Germany toward LawZero. While

LLM Gateway

Coverage

On September 11, 2026, Sakana AI launched Fugu Ultra v2.0 alongside Fugu Max v1.0, positioning both as orchestration engines rather than monolithic models. Fugu Ultra v2.0 is built on learned model coordination described in two ICLR 2026 papers: TRINITY, which uses an evolved LLM coordinator, and Conductor, which uses Fugu Ultra v2.0 is priced at $5 per million input tokens, $30 per million output tokens, and $0.50 per million cached input tokens, with a premium tier of $10/$45/$1.00 for contexts exceeding 272K tokens. On benchmarks, Fugu Ultra v2.0 achieves best or joint-best results on five of eight evaluations, including GDP.pdf,

LLM Gateway

CoverageBenchmark

According to a Prompt20 reasoning leaderboard entry dated September 19, 2026, Sakana AI's Fugu Ultra v2.0 is characterized as a learned orchestration system: a language model trained to route tasks across a pool of open-weight and specialist models. The leaderboard explicitly names Sakana AI as the creator and reports The same leaderboard row places Fugu Ultra v2.0 in the top tier of reasoning/math models as of mid-September 2026, sitting alongside entries like GLM 5.3 FlashX and DeepSeek V4.1 Flash. Because the benchmarks are Sakana-reported and not independently reproduced by the leaderboard, the figures should be treated as vendo