Fugu Max & Fugu Ultra v2 Review: Did Sakana Finally Make Multi-Agent Orchestration More Efficient Than Bigger Models?
Sakana Fugu is taking a different route from the rest of the AI market. Instead of trying to win by putting more compute into one monolithic model, Fugu is itself an orchestration model that decides how a pool of specialized language models should work together on a task. Fugu Max and Fugu Ultra v2 are the latest step in that strategy, released on September 11, 2026 with the same underlying orchestration architecture aimed at two different goals: lower cost and maximum capability.
Fugu Max is the cost-performance model. Sakana says it expands the pool to its largest collection of open and specialized models yet, including NVIDIA's Nemotron family through its NVIDIA collaboration. Its goal is to route each subtask to the leanest model that can solve it, then combine the results behind one API. Sakana reports best-overall results on six benchmarks and says Fugu Max expands the cost-performance Pareto frontier on seven of ten benchmarks.
Fugu Ultra v2 is the quality-first version. It keeps the same orchestration idea but targets harder multi-step reasoning, autonomous research and full-stack software engineering. Sakana reports best or joint-best results on five of eight benchmarks and top-two placement on seven of eight, including 74.3 on DeepSWE and 48.3 on Chartography. Sakana also states that Fable 5, Fable 5.1 and GPT-6-Astra are not in the Fugu Ultra v2 model pool.
QUICK ANSWER
Fugu Max and Fugu Ultra v2 are two operating points of the same Sakana Fugu multi-agent orchestration architecture. Fugu Max is designed to maximize cost-performance, while Fugu Ultra v2 is designed to maximize output quality on complex, multi-step tasks. Both are available through Sakana's OpenAI-compatible API.
Fugu Max is the more interesting release if your priority is economics. Sakana lists it at $2 per million input tokens and $6 per million output tokens and says it achieves the best overall score on Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench and SWEFish. Sakana also says it pushes the cost-performance frontier on seven of ten benchmarks.
Fugu Ultra v2 is the stronger choice for maximum quality. Sakana reports best or joint-best scores on GDP.pdf, Chartography, SWEFish, DeepSWE and Toolathon, with top-two results on seven of eight benchmarks. Its DeepSWE score is 74.3 and its Chartography score is 48.3.
My verdict: 9.2/10 for Fugu Max and 9.4/10 for Fugu Ultra v2. Fugu Max is the more compelling choice for high-volume enterprise agents and cost-sensitive production. Fugu Ultra v2 is the better option for research, difficult coding, complex reasoning and workloads where answer quality matters more than token cost.
1. What Is Sakana Fugu?
Sakana Fugu is a multi-agent system delivered through a model API. A Fugu request can dynamically involve multiple language models, with the system deciding which models should contribute and how they should communicate. Sakana's original research describes Fugu as a learned orchestration model rather than a manually programmed collection of fixed roles.
The research foundation comes from two ICLR 2026 directions, TRINITY and the Conductor. TRINITY uses a lightweight coordinator to orchestrate model roles, while the Conductor uses learned coordination strategies to construct agent-to-agent communication patterns. The commercial Fugu system builds on these ideas and exposes the result through one endpoint.
That distinction matters because Fugu is not simply a conventional model router. The orchestration logic is part of what Sakana trained the system to do. Fugu can select models, delegate subtasks, coordinate intermediate work and synthesize a final result without the user writing the full multi-agent workflow by hand.
2. What Changed With Fugu Max and Fugu Ultra v2?
The September 11 release makes the product strategy much clearer. Sakana is explicitly separating the frontier into two dimensions: capability and cost. Fugu Max moves toward the cost-efficient side by expanding the model pool and routing tasks to cheaper specialists. Fugu Ultra v2 moves upward by optimizing the same orchestration architecture for the hardest multi-step tasks.
The two versions therefore should not be treated as a conventional small-versus-large model choice. They are two different optimization targets built on the same orchestration idea.
3. Fugu Max: The Cost-Performance Play
Fugu Max is built around a simple idea: the best model for a task is often not the biggest model available. A request can contain several subtasks, and each subtask can be routed to a different model based on what it actually needs.
For example, an enterprise research workflow might use one specialist to extract structured facts, another to reason over conflicting evidence, and a smaller model to format the final result. Fugu Max's job is to coordinate those calls without forcing the developer to manage every provider directly.
Sakana says the expanded Fugu Max pool includes open-weight and specialized models, including NVIDIA's Nemotron family. The company reports best-overall results on six benchmarks and a Pareto-frontier improvement on seven of ten benchmarks.
4. Fugu Max Benchmarks
5. Fugu Ultra v2: The Quality-First Model
Fugu Ultra v2 takes the opposite side of the same architecture. Instead of minimizing cost, it uses the orchestration system to maximize performance on tasks that benefit from sustained reasoning, multiple specialists and iterative verification.
Sakana says Fugu Ultra v2 is aimed at complex multi-step reasoning, autonomous research and full-stack software development. The model's reported results place it in the top two on seven of eight cited benchmarks and best or joint-best on five.
6. DeepSWE: A Strong Coding Signal
Fugu Ultra v2's 74.3 DeepSWE score is one of the most important results in the release. DeepSWE is designed around real-world software engineering rather than isolated code generation, so performance on it is especially relevant to coding agents.
Sakana says Fugu Ultra v2 reached 74.3 without Fable 5, Fable 5.1 or GPT-6-Astra in its pool. This supports Fugu's broader argument that a learned orchestration layer can combine open and specialized models to produce strong results even when newer proprietary frontier systems are unavailable.
For developers, this does not mean every repository task will be better than a single frontier model. It means the model selection and coordination layer itself is becoming part of the performance equation.
7. Chartography: Where Fugu Ultra v2 Stands Out
Chartography is another standout result because it tests visual reasoning and data interpretation. Sakana reports 48.3 for Fugu Ultra v2, compared with 27.3 for Opus 5 and 29.5 for Fable 5 in the cited comparison.
This is the kind of task where orchestration can be useful. One agent can interpret the chart, another can challenge the interpretation, and another can construct the final reasoning. Fugu's learned coordination system is designed to discover useful collaboration patterns rather than relying on a single fixed chain.
8. Fugu Max vs Fugu Ultra v2
9. How Fugu Differs From a Normal Multi-Agent Framework
Traditional multi-agent systems require developers to explicitly define agents, prompts, tools, handoffs and often the topology of the workflow. Fugu moves more of that design into the learned orchestration layer.
That abstraction can dramatically reduce application complexity. The tradeoff is visibility: the developer gives up some direct control over every internal step in exchange for a simpler interface and learned orchestration.
10. API and Compatibility
Both Fugu Max and Fugu Ultra v2 are available through Sakana's standard OpenAI-compatible API. Sakana says existing Fugu users can switch to Max or Ultra v2 with a one-line model parameter change, without a broader migration.
The API supports standard model-client patterns, and Sakana's general Fugu documentation states that the platform can work with OpenAI-compatible clients and coding harnesses. This makes integration significantly easier than rebuilding a custom multi-agent system from scratch.
11. Pricing and Cost Efficiency
Fugu Max has a published launch price of $2 per million input tokens and $6 per million output tokens. Sakana says this output price is 40-60% lower than Sonnet 5, GPT 5.6 Terra and Kimi K3 in its comparison.
The current public pricing page still lists fixed Fugu Ultra v1.1/v1.0 rates at $5 per million input tokens and $30 per million output tokens, with a higher long-context rate above 272K tokens. The September 11 Fugu Ultra v2 announcement does not state a new v2 token price on its launch page, so this review does not invent a new Ultra v2 rate.
This is an important distinction for buyers. Fugu Max has an explicit new price. Ultra v2 has a new release but the public pricing page retrieved for this review still describes the previous Ultra versions. Check the live console before budgeting production use.
12. Supply-Chain Resilience and Model Independence
Sakana's model orchestration strategy has a second value beyond benchmarks: it reduces dependence on one model provider. Fugu can work with a swappable pool of models, and Sakana describes this as a way to reduce vendor lock-in, API revocation risk and geopolitical or supply interruptions.
This does not remove all dependency. The Fugu platform and Sakana's own orchestration service become part of the stack. It changes the dependency shape from one frontier provider to an orchestration platform managing a pool of interchangeable providers.
For enterprises, that can be valuable when continuity, procurement and model availability matter as much as raw benchmark scores.
13. Data, Privacy and Provider Control
Sakana's current Fugu documentation says standard Fugu users can opt specific models out of the pool through console settings to meet data, privacy and compliance needs. Fugu Ultra uses the full pool for its performance target and does not expose the same opt-out flexibility.
This is an important architectural tradeoff. Fugu's strength comes partly from the diversity of the pool, but that diversity can conflict with organizations that have restrictions on where sensitive data can be processed.
For regulated workloads, the right question is not just whether Fugu is accurate. It is which models can see the request, what data leaves the approved boundary, and what routing controls are available.
14. Best Use Cases
15. Limitations You Should Know
- Fugu Max and Ultra v2 are orchestration systems, so performance depends on the internal model pool and coordination strategy.
- Many headline benchmark results come from Sakana's evaluation program and should be read as vendor-reported results.
- SWEFish is an internal Sakana benchmark, so it is best used as a product-specific signal rather than a universal industry ranking.
- The September 11 release page does not publish a separate Fugu Ultra v2 token price; the existing pricing page still lists v1.x Ultra rates.
- Fugu Ultra uses the full pool for performance, which can reduce provider-level control.
- Orchestration adds system complexity behind the API even though integration is simpler.
- Model pools will change over time, so benchmark results and routing behavior can evolve.
16. Recommended Production Workflow
Use Fugu Max as the default worker for high-volume production tasks.
Escalate complex reasoning, autonomous research and difficult coding to Fugu Ultra v2.
Define explicit quality checks after important tool or agent steps.
Use structured output when Fugu results feed another service.
Audit the internal model pool for privacy and compliance requirements.
Track cost per completed task instead of token price alone.
Keep a direct fallback model available for critical paths if the orchestration service is unavailable.
17. How to Evaluate Fugu Max and Ultra v2 Yourself
Run both Fugu variants on the exact tasks your team cares about. Use the same documents, tools and success criteria you use in production.
For Fugu Max, focus on whether cost savings survive real traffic. For Ultra v2, focus on whether the additional orchestration depth produces better final outcomes on the hardest tasks.
18. Fugu Max vs Fugu Ultra v2: Which Should You Use?
Choose Fugu Max when your primary problem is inference economics. It is designed to find a capable model for each subtask and keep total token cost low. The $2 input and $6 output pricing makes the intended use case clear: production agents that need a lot of intelligence without paying frontier-model rates at every step.
Choose Fugu Ultra v2 when your primary problem is task difficulty. Its reported performance on DeepSWE, Chartography and the broader eight-benchmark set makes it the better fit when a task can justify additional orchestration and a higher quality target.
19. Is Fugu Max Worth It?
Fugu Max is worth testing for teams with meaningful AI-agent traffic. Its core advantage is not one benchmark score. It is the possibility of getting close to frontier quality without routing every subtask to an expensive model.
The six-benchmark best-overall claim and seven-of-ten Pareto-frontier result make that strategy credible. The $2/$6 token pricing then turns the benchmark story into a practical business case.
The key production question is routing quality. If Fugu consistently chooses the right lean model for each subtask, the savings should appear as lower cost per successful workflow. If it over-escalates to expensive agents, the advantage shrinks.
20. Is Fugu Ultra v2 Worth It?
Fugu Ultra v2 is worth testing for high-value tasks where correctness matters more than raw token cost. The DeepSWE and Chartography results are particularly compelling, and the top-two placement across seven of eight benchmarks suggests a broad quality advantage in the cited evaluation suite.
It is also strategically interesting because Sakana says Fugu Ultra v2 achieved those results without Fable 5, Fable 5.1 or GPT-6-Astra in its pool. The system demonstrates what a learned orchestration layer can do with a swappable collection of open and specialized models.
The tradeoff is that the current public pricing page does not yet publish a distinct v2 token rate, so production buyers should verify the live console pricing before committing budget.
21. Final Verdict
Fugu Max and Fugu Ultra v2 are important because they make orchestration itself the product. Sakana is no longer presenting multi-agent coordination as an implementation detail developers build on top of a model. It is delivering the orchestration engine as a model API.
Fugu Max is the stronger economic story. At $2 per million input tokens and $6 per million output tokens, Sakana is explicitly targeting the cost-performance frontier. Its reported wins on Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench and SWEFish give the strategy measurable support.
Fugu Ultra v2 is the stronger technical story. A 74.3 DeepSWE score, 48.3 Chartography result, five best-or-joint-best scores across eight benchmarks and top-two placement on seven of eight show what learned orchestration can do when maximum quality is the target.
The main limitation is benchmark provenance. These are Sakana-reported evaluation results, and SWEFish is an internal benchmark. Buyers should reproduce the comparison on their own workflows, especially where routing, latency, privacy and token economics affect the final outcome.
My rating: 9.2/10 for Fugu Max, 9.4/10 for Fugu Ultra v2 and 9.3/10 for the Fugu orchestration approach overall.
Bottom line: Fugu Max is the better default for production-scale agent workloads where cost matters. Fugu Ultra v2 is the better choice for complex reasoning and coding where quality matters most. The bigger lesson is that the competitive unit in AI may increasingly be an orchestrated system rather than a single model.
Frequently Asked Questions
What is Fugu Max?
Fugu Max is Sakana Fugu's cost-performance-oriented orchestration model, designed to route subtasks to the leanest capable model and combine the results.
What is Fugu Ultra v2?
Fugu Ultra v2 is the quality-first version of the same Fugu orchestration architecture, designed for complex reasoning, autonomous research and full-stack software engineering.
When were Fugu Max and Fugu Ultra v2 released?
Sakana announced both models on September 11, 2026.
How much does Fugu Max cost?
Sakana lists $2 per million input tokens and $6 per million output tokens.
How much does Fugu Ultra v2 cost?
The September 11 release page does not state a new v2 token price. Sakana's current public pricing page still lists v1.x Ultra at $5/M input and $30/M output, with higher long-context rates above 272K tokens.
What are the Fugu Max benchmark results?
Sakana reports best overall scores on Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench and SWEFish.
What are the Fugu Ultra v2 benchmark results?
Sakana reports best or joint-best scores on GDP.pdf, Chartography, SWEFish, DeepSWE and Toolathon, with top-two placement on seven of eight benchmarks.
Does Fugu Ultra v2 use Fable 5 or GPT-6-Astra?
Sakana explicitly says those models are not in the Fugu Ultra v2 pool.
Does Fugu use multiple models behind one API?
Yes. Fugu dynamically orchestrates a pool of specialized and frontier models behind a single OpenAI-compatible API.
Can I use Fugu with existing OpenAI-compatible tools?
Yes. Sakana provides an OpenAI-compatible API and says existing clients and coding harnesses can point to the Fugu endpoint.
Is Fugu open source?
The commercial Fugu service is a hosted orchestration product. Sakana publishes its research and technical work, but the commercial service is not the same thing as downloading a single open-weight model.
Which should I use, Fugu Max or Fugu Ultra v2?
Use Fugu Max for cost-sensitive production and high-volume agents. Use Fugu Ultra v2 for the hardest research, coding and reasoning tasks.
Recommended Blogs
Meta Muse Spark 1.3 Review: Coding, Price & Is It Worth It? (2026)
Gemini 3.8 Flash Review: Accuracy, Price & Is It Worth It? (2026)
Quasar 438B Review: Benchmarks, Speed, Price & Is It Worth It? (2026)
Qwen 3.8 Max 0902 Review: Benchmarks, Price & Is It Worth It? (2026)
Mercury 2.5 AI Model Review: Speed, Price & Is It Worth It? (2026)
How to Secure AI Coding Agents: Permissions, Sandboxing, MCP & Secrets
Resources & Community
Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications. Whether you are a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.
Agentic AI Launchpad 2026
A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews and a builder community network.
Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026
Free AI Resources
Access free tools, workshops and micro-learning to keep building.
