Summary
On September 11, 2026, Tokyo-based Sakana AI released Fugu Ultra v2 and a cheaper sibling, Fugu Max. Rather than a single monolithic model, Fugu is a learned orchestrator: one OpenAI-compatible API request is routed behind the scenes across a pool of frontier open-weight and specialized models. Fugu Ultra v2 ships a 1,000,000-token context at $5 input / $30 output per million tokens and posts top or joint-top scores on five of eight of Sakana's headline agentic benchmarks, including 74.3 on DeepSWE — edging GPT-6 Astra (74.1) and Claude Fable 5.1 (67.4) despite excluding both from its pool.
What changed
Sakana AI launched Fugu Ultra v2 and Fugu Max on September 11, 2026 — orchestration models delivered through a single OpenAI-compatible API that dynamically route each request across a proprietary pool of models, with a 1M-token context and $5/$30 per-million-token pricing on Ultra v2.
Why it matters
Fugu is a bet that learned orchestration across many models beats any single frontier model on agentic tasks at lower cost — a direct challenge to the one-big-model strategy of OpenAI and Anthropic, and validation for the routing layer that gateways and coding tools are racing to build.
Evidence excerpt
Fugu is not a single model; it is an orchestrator trained to hand each task to other models and stitch the answers back together. Fugu Ultra v2 scored 74.3 on DeepSWE, edging GPT-6 Astra's 74.1 and Claude Fable 5.1's 67.4, at $5/$30 per 1M tokens with a 1M-token context.