# Sakana Fugu Max: Orchestration at $2/M Tokens > Source: > Published: 2026-09-12 11:10:31+00:00 Sakana AI shipped two new Fugu models on September 10–11 — and neither is a model in the conventional sense. Fugu Max and Fugu Ultra v2 are orchestration engines: systems trained to route your request across a pool of AI models behind a single OpenAI-compatible endpoint. Fugu Max is the cost-performance tier, priced at $2 per million input tokens and $6 per million output tokens — 40 to 60 percent below Sonnet 5 and GPT-5.6 Terra on output cost. ## The Architecture Behind the Price Tag Fugu isn’t running your requests through one large model. It uses TRINITY, a lightweight coordinator (roughly 0.6 billion parameters) that assigns roles — Thinker, Worker, or Verifier — to a pool of larger specialized models per request. On top of that sits the Conductor: a 7B model trained with reinforcement learning to design the natural-language coordination strategy on the fly. The Conductor can call itself recursively, scaling compute at test time without any API change on your end. Fugu Max expands the model pool significantly, adding the [NVIDIA Nemotron family through a new Sakana-NVIDIA collaboration](https://www.techtimes.com/articles/320781/20260716/sakana-fugu-adds-nemotron-prove-open-models-can-route-past-proprietary-walls.htm). The routing logic is learned, not hand-coded — meaning the orchestrator was trained to generalize across coding, math, reasoning, and knowledge tasks rather than following static rules you’d maintain yourself. ## The Cost Argument, Spelled Out | Model | Input ($/M) | Output ($/M) | |---|---|---| | Fugu Max | $2.00 | $6.00 | | Fugu Ultra v2 | $5.00 | $30.00 | | Claude Sonnet 5 | ~$3.00 | ~$15.00 | | GPT-5.6 Terra | ~$2.50 | ~$15.00 | At 100 million output tokens per month — a realistic volume for a production coding tool or content pipeline — Fugu Max costs $600 versus $1,500 for Sonnet 5. That $900 monthly gap is the actual argument here. Sakana reports Pareto frontier improvements on 7 of 10 internal benchmarks per their [official announcement](https://sakana.ai/fugu-max-release/), though independent validation hasn’t arrived yet. ## What Fugu Ultra v2 Adds Fugu Ultra v2 keeps the 1-million-token context window but ships with three additions that weren’t in v1: native visual input (images and PDFs), function calling via `tools` and `tool_choice`, and structured outputs through a JSON schema in `response_format`. The training cutoff moved to August 28, 2026 — the Conductor was re-tuned against current frontier models. This tier costs $5 input / $30 output and targets agentic workflows where you need reasoning, tool calls, and visual understanding in a single request. ## Migration Is One Line If you’re already calling Fugu, upgrading to Fugu Max or Ultra v2 is a one-parameter change: ``` # Before model="fugu-ultra" # After -- same API key, same client, same request shape model="fugu-max" ``` No library swap. No new authentication. No rethinking your integration. For a more detailed breakdown of pricing across providers, [OpenRouter’s Fugu Ultra v2 listing](https://openrouter.ai/sakana/fugu-ultra-v2) tracks live rates and availability. ## The Tradeoffs You’re Accepting Fugu Max is not a universal replacement for calling frontier models directly. The constraints are real: - **No EU/EEA access.** Sakana is working on GDPR compliance but has given no timeline. If your infrastructure is in Europe, Fugu is not an option today. - **No open weights.** Hosted API only. You cannot self-host, audit the weight stack, or run offline inference. - **Latency uncertainty.** Orchestration systems coordinate multiple models per request. In June 2026, users reported 30-minute waits at peak load on earlier Fugu versions. Sakana has not addressed this directly in the Fugu Max notes. - **Opaque routing.** You don’t control or see which model in the pool handled your request — a concern for reproducibility and debugging in regulated environments. For teams needing open weights, EU data residency, or deterministic latency, Fugu Max is the wrong tool. For teams running output-heavy workloads with existing Fugu integrations and no EU requirements, the cost reduction is concrete. The [Sakana Fugu technical report](https://arxiv.org/html/2606.21228v1) covers the TRINITY architecture in depth if you want to evaluate the routing claims before committing. ## Context: What This Is Competing Against Fugu Max arrives weeks before OpenAI’s DevDay on September 29, where Managed Agents — OpenAI’s own platform for orchestrated multi-step agent workflows — are expected to launch publicly. Sakana’s bet is that developers want orchestration delivered as a model endpoint, not a platform to configure. The difference in abstraction level is real: Fugu requires no pipeline code, while Managed Agents will presumably require setup and configuration. Whether that tradeoff wins depends on how much routing control developers actually want. We covered the [original Conductor architecture](https://byteiota.com/sakana-fugu-rl-conductor/) when Sakana first introduced RL-trained coordination in May — the Fugu Max release is that system, scaled and repriced. Both models are live now through Sakana’s OpenAI-compatible API.