# Sakana AI Ships Fugu Max and Ultra v2, Orchestrators Instead of Models

Page: https://hirantha.me/blog/sakana-ai-ships-fugu-max-and-ultra-v2-orchestrators-instead-of-models
Published: 2026-09-12
Written by: hirantha.me Newsroom (AI-generated). An AI model wrote this from the sources listed below; how it is made: https://hirantha.me/about/newsroom
Topics: sakana ai, ai agents, llm routing, multi-agent systems, ai pricing

> Sakana AI's new Fugu Max and Ultra v2 route each task to other AI models rather than answering it directly, and Ultra v2 leaves Claude and GPT-6 out entirely.

## Two releases that don't answer questions themselves

Sakana AI released two new systems this week, Fugu Max and Fugu Ultra v2, and neither one is a foundation model in the usual sense. Rather than being trained to answer a prompt directly, both are what the company calls learned orchestrators: a routing layer trained to look at an incoming task and hand it off to whichever model in a pool of other systems is best suited to solve it, calling on itself recursively as needed.

The two variants are tuned for opposite goals. Fugu Max asks what the best result is that can be delivered at the lowest cost, routing each task to the smallest model in its pool likely to handle it. Fugu Ultra v2 asks the opposite question, what the highest capability achievable on a hard, multi-step task looks like, and routes for quality even if that costs more tokens. Both went live this week through an application programming interface (API) that is compatible with OpenAI's API format, so existing tooling built for that format needs no rewrite to point at Fugu instead.

## Built on Sakana's own routing research

The orchestration layer draws on two papers Sakana presented at the 2026 International Conference on Learning Representations (ICLR), a major AI research venue. One, called TRINITY, assigns each turn of a task one of three roles: Thinker, Worker or Verifier. The other, called The Conductor, is a routing policy trained with reinforcement learning specifically to discover how to coordinate across multiple models rather than to answer questions itself. Together they form the mechanism deciding, task by task, which model in the pool actually does the work.

## Pricing and what's in the pool

Fugu Max is priced at $2 per million input tokens and $6 per million output tokens, which Sakana says lands 40 to 60 percent below comparable offerings from Anthropic's Claude Sonnet 5, OpenAI's GPT-5.6 Terra and Moonshot AI's Kimi K3. Sakana has not published separate list pricing for Fugu Ultra v2, since its routing optimizes for outcome quality rather than a fixed per-task cost target.

The pool both variants draw from is made up of open-weight and specialized models, including Nvidia's Nemotron family, and Sakana describes the pool as swappable: models can be added or removed from behind the API without an application built on top having to change anything. The stated reason is to reduce exposure to any single model provider's price changes, access revocations or shutdowns.

The most pointed decision in the release is what Sakana deliberately left out. Fugu Ultra v2's evaluation and routing pool exclude Anthropic's Claude Fable 5 and Fable 5.1 and OpenAI's GPT-6 Astra entirely. Sakana frames this as proof the system does not need access to a proprietary frontier model to post competitive scores, and as a hedge against the same vendor lock-in it says the swappable pool design is meant to avoid.

## The numbers Sakana is citing

On benchmark results Sakana has published, Fugu Max posts the best overall score among the systems it compared against on six separate benchmarks, including GPQA Diamond, a graduate-level science question-answering benchmark, and AutomationBench. Fugu Ultra v2 comes out best or tied for best on five of eight benchmarks tested, among them a visual-reasoning benchmark called Chartography, where it scored 48.3 against 27.3 for Anthropic's Claude Opus 5 and 29.5 for Claude Fable 5, and a software-engineering benchmark called DeepSWE, where its score of 74.3 is one Sakana says beats models that cost three to five times as much per token to run. These figures come from Sakana's own release materials rather than an independent leaderboard, so they are worth treating as a vendor's best case rather than a neutral ranking.

Fugu Max and Ultra v2 are reachable only through Sakana's hosted API; there are no downloadable weights for self-hosting, and the service is not available to users in the European Union or European Economic Area.

## What it means in practice

The interesting bet here is less about raw benchmark scores and more about the architecture: instead of building one more frontier model to compete with Claude or GPT-6 head-on, Sakana is betting that a well-trained router sitting on top of a pool of cheaper, swappable models can match or beat them on cost and, on some tasks, on quality. For a team building an application today, that trade-off is worth weighing directly against a single-model API: Fugu's pricing and swappable pool reduce exposure to any one vendor's price or policy changes, but the published results are self-reported, so anyone evaluating it for production use should run it against their own workload rather than take the benchmark comparisons at face value.

## Questions

**What makes Fugu Max and Ultra v2 different from a typical AI model?**
Neither answers a prompt directly. Both are routing systems trained to send each task to whichever model in an underlying pool is best suited to solve it, rather than being a single model that generates the answer itself.

**Is Fugu Max actually cheaper than Claude or GPT?**
Sakana's own pricing comparison puts Fugu Max 40 to 60 percent below Claude Sonnet 5, GPT-5.6 Terra and Kimi K3 on a per-token basis, though these are the vendor's own figures rather than an independent benchmark.

**Why does Fugu Ultra v2 leave out Claude and GPT-6 Astra?**
Sakana says this is deliberate, to demonstrate the system can post competitive scores without relying on a proprietary frontier model, and to avoid the vendor lock-in and access-revocation risk that comes with depending on one provider.

**Can I download and self-host Fugu's models?**
No. Fugu Max and Ultra v2 are only accessible through Sakana's hosted, OpenAI-compatible API; there are no released weights for self-hosting.

**Is Fugu available in the European Union?**
No, Sakana's release does not make the service available to users in the European Union or European Economic Area.

## Sources

- [Introducing Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier](https://sakana.ai/fugu-max-release/), Sakana AI
- [Sakana AI Launches Fugu Max and Fugu Ultra v2 for Cheaper, Stronger Multi-Agent Orchestration](https://www.marktechpost.com/2026/09/10/sakana-ai-launches-fugu-max-and-fugu-ultra-v2-for-cheaper-stronger-multi-agent-orchestration/), MarkTechPost
- [Sakana AI Splits Fugu Into Max and Ultra v2 to Cut Costs 60%](https://alphasignal.ai/news/sakana-ai-splits-fugu-into-max-and-ultra-v2-to-cut-costs-60), AlphaSignal