Articles tagged "Sakana AI"

**Sakana AI officially launched Sakana Fugu on June 22, 2026** ([fugu-release](https://sakana.ai/fugu-release/) / [product page](https://sakana.ai/fugu/) / [gihyo.jp](https://gihyo.jp/article/2026/06/sakana-fugu) / [GIGAZINE](https://gigazine.net/gsc_news/en/20260622-sakana-fugu-multi-agent-system-ai)). Critically, this is **not a next-generation Japanese LLM — it is an LLM trained to call other LLMs**, a 'conductor' model that dynamically orchestrates frontier models inside the loop. When you send a query, Fugu itself either (1) answers directly when it can, or (2) for complex multi-step tasks **selects, dispatches, verifies, and integrates** from an agent pool that includes **GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro** and others. **Academic basis**: two ICLR 2026 papers — **TRINITY** (an evolutionarily optimized LLM coordinator that dynamically assigns Thinker / Worker / Verifier roles) and **Conductor** (RL-discovered coordination strategies expressed in natural language). **Two variants**: **Fugu** (everyday tasks, low latency) and **Fugu Ultra** (hardest problems, deep coordination — pool composition is fixed and cannot be excluded). **Benchmarks**: **SWE-Bench Pro 73.7** (reported to beat Claude Opus 4.8, per XenoSpectrum), Terminal-Bench 2.1 above Anthropic's latest, Charxiv Reasoning above Claude Mythos Preview — but **lags on Humanity's Last Exam (HLE)**. Sakana's own framing is conservative: "shoulder-to-shoulder with Fable 5 and Mythos Preview," not blanket dominance. **Pricing**: Fugu Ultra at **$5/M input ($10/M >272K) and $30/M output ($45/M >272K)**, plus **subscriptions at Standard $20 / Pro $100 / Max $200 per month** (both Fugu and Fugu Ultra). Enterprise is usage-based. **OpenAI-compatible API** at console.sakana.ai. **Not available in the EU/EEA** pending GDPR compliance; Japan-region usage works. **The strategic point is structural resilience, not raw performance** — escape from single-vendor dependence and diversification against export-control risk (directly continuing our Sakana Marlin column's Fable 5 export-restriction thread). BuildFastWithAI calls it 'the orchestration model that routes around export controls,' and Clanker Cloud frames it as 'Model Orchestration Is Becoming the Product.' **Fugu's own parameter count, Japanese-specific benchmark scores (ELYZA / JMMLU / JMT-Bench), and individual statements from David Ha / Llion Jones are not yet confirmed**, leaving 'thin wrapper over external APIs' criticism and independent verification as open questions.

Sakana AI's first commercial product 'Marlin,' launched June 15, 2026, is an autonomous research agent — not an LLM. Combining AB-MCTS (Adaptive Branching Monte Carlo Tree Search) with multi-LLM collaboration across OpenAI o4-mini, Google Gemini 2.5 Pro, and DeepSeek R1-0528, Marlin operates autonomously for up to ~8 hours per task to generate tens-to-100+ page reports and executive slides. Designed for financial institutions, corporate planning, consulting, and think tanks, it differs fundamentally from OpenAI Deep Research and Gemini Deep Research in both purpose and architecture. This guide covers everything from its technical design to pricing, competitor comparison, and what it means for Japanese enterprises.