Sakana AI Launches Fugu Ultra v2, a Multi-Agent Orchestrator With 1M-Token Context

Sakana AI has released Fugu Ultra v2, the higher-performance model in its Fugu family, now available via OpenRouter. The system is built as a learned multi-agent orchestration architecture: a language model trained to route tasks across a fixed pool of open and specialized models, and to recursively call instances of itself, rather than functioning as a single trained checkpoint.

Specifications

Fugu Ultra v2 ships with a 1.0 million token context window and a knowledge cutoff of August 2026. It was released on September 11, 2026, according to OpenRouter's model listing.

Pricing is set at $5 per 1M input tokens and $30 per 1M output tokens, with cached input reads priced at $0.50 per 1M tokens. Prompts exceeding 272,000 tokens are billed at a higher rate, and orchestration tokens — the tokens the system consumes internally while routing and coordinating sub-agent calls — are billed as standard input/output tokens.

Measured P50 latency is 9.21 seconds, with throughput around 15 tokens per second, per OpenRouter's live performance data.

Capabilities

According to Sakana AI, Fugu Ultra v2 is designed to prioritize answer quality over speed on complex multi-step reasoning, autonomous research tasks, and full-stack software development work. The model supports:

• Configurable reasoning effort levels: high, xhigh, and max
• Function calling and structured outputs
• Image and PDF input (multimodal)
• Built-in web search

Sakana AI states the model "does not rely on individual proprietary frontier models in its pool," implying the orchestration system routes exclusively to open or Sakana-specialized models rather than calling out to systems like GPT or Claude. This claim has not been independently verified, and no third-party benchmark scores (MMLU, HumanEval, or similar) have been published alongside the release.

What this means

Fugu Ultra v2 represents a bet on orchestration over scale: instead of training one larger monolithic model, Sakana AI is packaging a router that dispatches work across existing open models and calls itself recursively for multi-step tasks. This approach could make the system cheaper to iterate on than training frontier-scale weights from scratch, but it also means end-to-end quality depends heavily on the strength of the underlying model pool and the router's decision-making — neither of which is disclosed in detail here.

The pricing structure is notable: at $5/$30 per 1M tokens, Fugu Ultra v2 sits above budget open-weight offerings but below top-tier proprietary reasoning models. The 272K-token pricing cliff and separate billing for orchestration tokens suggest real compute costs from the recursive self-calling architecture, which buyers should factor into usage estimates for long-context or multi-agent workloads.

Without independent benchmark data, it's not yet possible to verify whether Fugu Ultra v2's routing approach outperforms single-model alternatives at comparable price points. Developers evaluating it for production use should test latency-sensitive workloads carefully — a 9.21-second P50 latency and 15 tokens/second throughput are markedly slower than typical single-model inference, a likely tradeoff of the multi-agent coordination overhead.