Sakana AI's Fugu Ultra v2 Beats GPT-6 and Claude

On September 11, 2026, Sakana AI shipped two new models — Fugu Ultra v2 and Fugu Max — alongside a provocation that should make the AI industry pause: the new flagship version no longer uses Claude Fable 5, Claude Fable 5.1, or GPT-6 Astra in its orchestration pool, yet it claims to beat all three on five of eight published benchmarks.

What Fugu Actually Is

Fugu is not a foundation model. It is a learned orchestrator — a system that routes each task across a pool of specialized models behind a single OpenAI-compatible API endpoint. The architecture is grounded in two papers from ICLR 2026: TRINITY and the Conductor, which demonstrated that a system trained to coordinate expert agents can outperform any single agent in its pool, provided the routing itself is learned rather than hand-coded.

When you send Fugu Ultra v2 a request, it decomposes the task, dispatches sub-problems to whichever pool members it has learned are best suited, and synthesizes a single response. The user sees none of this: one endpoint, one bill, no model-management overhead.

The Benchmark Claims

According to Sakana's internal evaluation, Fugu Ultra v2 is best or joint-best on five of eight tasks. The most striking numbers: on Chartography it scores 48.3 against Claude Opus 5's 27.3 and Claude Fable 5's 29.5. On DeepSWE, a software-engineering benchmark, it posts 74.3.

A word of caution: these are Sakana's own numbers, and no independent reproduction exists at time of writing — there is no Artificial Analysis page for any Fugu model. Treat them as a strong prior, not settled fact.

The Pool Swap

The most notable decision Sakana made was which models to remove. Fugu Ultra v1 leaned on Anthropic's and OpenAI's strongest models to fill the top of its quality curve. v2 drops them entirely. That is partly a cost play — the split into Max and Ultra v2 cuts total API spend by roughly 60% for typical workloads — but it is also an architectural confidence claim: the orchestration layer is now where the value lives, not in which model anchors the pool.

Fugu Max: The Other Half

Fugu Max, launched alongside Ultra v2, trades raw quality for significantly lower cost, using a shallower agent pool tuned for speed and economy. For teams running high-volume, lower-stakes workloads, it fills a price-performance gap that neither GPT-6 Astra nor Claude Opus 5 address.

Pricing

Fugu Ultra v2 is priced at $5 per million input tokens and $30 per million output tokens, with $0.50 for cached input reads — rising to $10 / $45 / $1.00 for contexts above 272K tokens. Both models are available now via Krater.ai and OpenRouter.

Whether the benchmark claims hold under independent scrutiny will determine whether Fugu Ultra v2 is a landmark or a well-marketed oddity. Either way, the architecture is genuinely different. The bet that orchestration beats scale is one worth watching.