What is the difference between Fugu Max and Fugu Ultra v2?

Both were released by Sakana AI on September 11, 2026 and both are 'multi-agent systems delivered as one model' — a learned orchestrator that routes each task across a pool of underlying models. Fugu Max v1.0 is the cost-performance tier: $2 per million input tokens and $6 output, flat regardless of context, orchestrating Sakana's largest pool including NVIDIA's Nemotron family. Fugu Ultra v2.0 is the quality tier: $5/$30 (rising to $10/$45 above 272K tokens of context), coordinating a deeper pool of expert agents for hard, high-stakes problems.

How does Fugu Max pricing compare with Claude Sonnet 5?

Fugu Max costs $2 input / $6 output per million tokens with $0.25 cached input; Claude Sonnet 5 costs $2 / $10 with $0.20 cache hits (Anthropic made the $2/$10 rate permanent in August 2026). Input prices are identical, but Fugu Max's output is 40% cheaper than Sonnet 5, 50% cheaper than GPT-5.6 Terra ($2/$12) and 60% cheaper than Kimi K3 ($3/$15). Fugu also bills web_search and web_fetch at $0.007 per call.

Which frontier models are inside Fugu's pool?

Sakana does not publish the full pool, but it states that Fugu Ultra v2.0's training cutoff is August 28, 2026 and that Claude Fable 5, Fable 5.1 and GPT-6 Astra are NOT in its pool. Fugu Max integrates a large number of open-weight and specialized models, including NVIDIA Nemotron through a Sakana–NVIDIA collaboration. Customers can opt specific providers or models out of the pool for data, privacy or compliance reasons. Billing never stacks: when multiple agents run, you pay a single rate based on the top-tier model involved.

Is Sakana Fugu available in the EU?

No. As of September 2026 Sakana states that Fugu is 'not yet available in the EU/EEA' while it works toward GDPR compliance. It is available through Sakana's OpenAI-compatible API, via OpenRouter and LLM Gateway, and on subscription plans at $20, $100 and $200 per month that include Fugu, Fugu Ultra and Fugu Max. Fugu Cyber, the security-specialised variant, is sales-only.

Quick Answer

Fugu Max vs Fugu Ultra v2 vs Claude Sonnet 5: Which to Pick

Published: September 13, 2026

The short answer

	Fugu Max v1.0	Fugu Ultra v2.0	Claude Sonnet 5
Maker	Sakana AI (Tokyo)	Sakana AI	Anthropic
Released	September 11, 2026	September 11, 2026 (v2.0)	GA June 30, 2026
What it is	Learned orchestrator over Sakana’s largest model pool, tuned for cost-performance	Learned orchestrator over a deeper expert pool, tuned for quality	Single frontier model
Input / output (per MTok)	$2 / $6, flat at any context	$5 / $30 (≤272K) · $10 / $45 (>272K)	$2 / $10
Cached input	$0.25	$0.50 (≤272K) · $1.00 (>272K)	$0.20
Extras	web_search / web_fetch $0.007 per call	—	Tools billed as tokens
Benchmark claim	Best overall on Terminal-Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench, SWEFish	Best or joint-best on 5 of 8 (GDP.pdf, Chartography, DeepSWE, Toolathon, SWEFish); top-2 on 7 of 8	Anthropic’s mid-tier workhorse; strong on coding, 1M context
Frontier models in pool	Not Fable 5 / 5.1 or GPT-6 Astra	Same exclusion; cutoff Aug 28, 2026	n/a
EU/EEA availability	No	No	Yes
Best for	High-volume coding, review, research where output tokens dominate cost	Kaggle-grade problems, paper reproduction, security analysis, patent research	Predictable single-vendor behaviour, EU users, Anthropic ecosystem

Pick Fugu Max if output tokens dominate your bill and you are outside the EU; pick Fugu Ultra v2 when a hard problem justifies 5x the output price; pick Claude Sonnet 5 when you need one accountable model, EU availability, or Anthropic-native tooling.

What Sakana shipped on September 11, 2026

Sakana AI released Fugu Max v1.0 and Fugu Ultra v2.0 on September 11, 2026, expanding the Fugu line to four models — Fugu, Fugu Ultra, Fugu Max and Fugu Cyber — all behind one OpenAI-compatible API. Fugu is not a single network. It is a learned coordinator (grounded in Sakana’s two ICLR 2026 papers, TRINITY and the Conductor) that assembles Thinker/Worker/Verifier roles from a pool of open and specialised models and routes each step of a task to whichever agent it predicts will do it best. Sakana’s pitch is “frontier-level performance without single-vendor dependency.”

Pricing is deliberately simple. Sakana says it “never stacks model fees”: when multiple agents are active you pay a single rate based on the top-tier model involved, and for Fugu Max that rate is a flat $2 input / $6 output per million tokens regardless of context length. Fugu Ultra v2.0 is $5 / $30 up to 272K tokens and $10 / $45 above it, with cached input at $0.50 / $1.00.

Fugu Max: the cost-performance play

Fugu Max orchestrates Sakana’s largest pool to date, integrating “an unprecedented number of open-weights and specialized models” including NVIDIA’s Nemotron family via a Sakana–NVIDIA collaboration. Sakana’s claims for it:

Best overall score on six benchmarks: Terminal-Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench and SWEFish (Sakana’s internal coding benchmark).
Output pricing 40–60% lower than Claude Sonnet 5, GPT-5.6 Terra and Kimi K3.
Expands the cost-performance Pareto frontier on seven of ten benchmarks.

Do the arithmetic on a code-review workload of 1M input and 1M output tokens per day: Fugu Max is $8/day, Sonnet 5 $12, GPT-5.6 Terra $14, Kimi K3 $18. The delta is entirely output price — inputs are identical to Sonnet 5 and Terra.

Fugu Ultra v2.0: the quality tier

Fugu Ultra coordinates a deeper pool of expert agents for “hard, high-stakes problems”; early users cite Kaggle competitions, paper reproduction, cybersecurity analysis and literature/patent investigations. Sakana claims v2.0 is best or joint-best on five of eight benchmarks (GDP.pdf, Chartography, DeepSWE, Toolathon, SWEFish) and top-2 on seven of eight. Its published demo is a mechanical CAD iris where Fugu Ultra’s blades actually close the aperture and competitors’ do not.

The most important disclosure is in the footnote: Fugu Ultra’s training cutoff is August 28, 2026, and Claude Fable 5, Fable 5.1 and GPT-6 Astra are not in its pool. Sakana frames this as a feature — Fugu “does not rely on individual proprietary frontier models to deliver frontier output” and therefore protects users from “vendor lock-in, API revocations, geopolitical turbulence, and sudden service cutoffs.” It also means Ultra’s ceiling is set by the open and specialised models it can reach, not by the newest closed flagships.

Claude Sonnet 5: the single-model baseline

Anthropic’s Claude Sonnet 5 went GA on June 30, 2026 at $2 / $10 per MTok, and Anthropic cancelled the scheduled September 1 increase to $3/$15 in August 2026, making $2/$10 the standard price. Cache hits cost $0.20, five-minute cache writes $2.50 and one-hour writes $4. What Sonnet 5 offers that Fugu does not:

One model, one behaviour. No orchestration layer choosing a different sub-model on Tuesday than it did on Monday; easier evals, easier prompt caching, easier incident review.
EU/EEA availability and Anthropic’s enterprise compliance surface (zero-data-retention options, Bedrock/Vertex/Foundry deployment).
Anthropic-native tooling: Claude Code, the Agent SDK, MCP-first integrations.

What it gives up: a higher output price than Fugu Max and no ability to swap providers when Anthropic changes a price, deprecates a model or tightens a safeguard.

Where the orchestration model is risky

Determinism and evals. An orchestrator that “dynamically assembles agents” is harder to regression-test than one model. Pin a model ID, snapshot your eval suite, and re-run it whenever Sakana announces a pool change.
Data residency. Your prompt may be forwarded to any model in the pool. Sakana lets you opt providers out for compliance — use it, and ask which vendors remain.
Benchmark provenance. All Fugu numbers are Sakana-run. SWEFish is internal. Independent leaderboards had not yet scored v2.0 or Max at publication.
Geography. No EU/EEA service as of September 2026.

Pricing plans beyond the API

Sakana also sells subscriptions for hands-on use — Standard $20/month, Pro $100/month, Max $200/month — each including Fugu, Fugu Ultra and Fugu Max with usage allowances; Sakana recommends pay-as-you-go tokens for “serious workloads.” Fugu Cyber (86.9% on Sakana’s cited security benchmark) is sales-only. Fugu is also listed on OpenRouter and LLM Gateway as of September 11, 2026.