Sakana AI Launches Fugu Max, a Multi-Agent Orchestration Model at $2/$6 per Million Tokens

Sakana AI has released Fugu Max, the cost-performance model in its new Fugu family, according to a listing on OpenRouter. Rather than a single trained network answering every query directly, Fugu Max is built as a learned multi-agent orchestration system: a language model trained to route tasks across a fixed pool of open-weights and specialized models — including NVIDIA's Nemotron family — and to recursively call instances of itself.

What Fugu Max does

According to Sakana AI, Fugu Max dynamically selects combinations of "expert" agents at inference time, aiming to improve both output quality and cost simultaneously rather than trading one for the other. The system supports configurable reasoning effort across three levels — high, xhigh, and max — giving users control over how much computation the orchestrator spends before returning an answer.

The model also supports function calling, structured outputs, image and PDF input, and built-in web search and web fetch tools. Notably, the internal "orchestration tokens" the system consumes while routing between agents are billed as standard input/output tokens, and pricing stays flat regardless of context length.

Specs and pricing

• Context window: 1,000,000 tokens
• Pricing: $2 per 1M input tokens / $6 per 1M output tokens
• Released: September 11, 2026
• Modalities: text, image, and PDF input; text output
• Reasoning modes: high, xhigh, max

Benchmark scores were not disclosed in the release materials, and Sakana AI has not published a technical paper or evaluation suite alongside the launch. Independent verification of the orchestration system's claimed cost-quality gains is not yet available.

What this means

Fugu Max is a bet on architecture over scale. Instead of training one larger model, Sakana AI is packaging a router that delegates work to existing open-weights systems, including NVIDIA's Nemotron line, and to copies of itself. If the routing logic works as described, it could let Sakana ship competitive quality without owning or training every underlying expert model — a meaningfully different cost structure than a typical frontier lab.

The flat, context-independent pricing at $2/$6 per million tokens is aggressive relative to many long-context models that charge premiums past certain token thresholds. But the architecture also introduces a new variable buyers haven't had to model before: orchestration overhead. Because routing and recursive self-calls are billed as standard tokens, actual cost-per-task will depend heavily on how many internal hops a given query triggers — something that won't show up in a simple per-token rate card.

Without published benchmarks, it's not yet possible to independently confirm Sakana's cost-performance claims relative to single-model competitors. The 1M token context window and multimodal input support put it in range of current frontier offerings on paper, but the real test will be third-party evaluations of latency, output quality, and effective per-task cost once developers start running it in production.