top of page

The Orchestration Revolution: How Sakana Fugu Max Is Challenging AI’s Most Expensive Models

Writer: Miao Zhang
Miao Zhang
3 days ago
10 min read
The artificial intelligence industry has spent much of the past decade competing to build increasingly capable foundation models. Larger parameter counts, greater training budgets, longer context windows, and more sophisticated reasoning systems have driven successive waves of progress. But the economics of AI are beginning to expose a different problem. For many real-world workloads, the most capable model available is not necessarily the most economically intelligent choice. A simple classification, data lookup, code transformation, or narrowly defined reasoning task may not justify consuming the computational resources of an expensive frontier model. The emerging question is therefore shifting from which model is smartest to which combination of models can solve a particular problem most efficiently. Sakana AI's latest Fugu releases illustrate this transition. With Fugu Max and Fugu Ultra v2, Sakana is advancing an orchestration model in which multiple open-weight and specialized AI systems are dynamically coordinated rather than relying on one monolithic model for every task. The strategy challenges conventional assumptions about AI pricing, vendor dependence, model specialization, and where economic value will ultimately accumulate across the AI stack. The implications extend well beyond one product release. If orchestration can consistently combine lower-cost models into systems capable of competing with more expensive frontier offerings, the competitive battlefield of AI may increasingly move from individual models to the intelligent infrastructure connecting them. The AI Industry Is Moving Beyond the Single-Model Paradigm The traditional foundation-model business has been relatively straightforward. A provider develops a powerful model, operates the infrastructure required to serve it, exposes its capabilities through an API, and charges customers according to usage. Developers then build applications around that model. This approach has obvious advantages. A single model provides a predictable interface, centralized optimization, and a relatively simple developer experience. Its weakness is economic inefficiency. Not every task requires the same level of intelligence. A complex software-engineering problem may demand substantial reasoning, while a straightforward extraction task may require only a smaller specialized model. Treating both workloads identically can result in unnecessary inference expenditure. Orchestration introduces a different architecture. Instead of asking one model to perform every operation, an orchestration layer can analyze a task, select appropriate models, divide work among agents, compare outputs, and coordinate the final result. The fundamental architecture becomes: Task → Routing → Specialized Models → Coordination → Verification → Final Output That creates a new optimization target, balancing capability against cost. Fugu Max Targets the Cost-Performance Frontier Sakana's Fugu Max is designed around this principle. The system orchestrates a large collection of open-weight and specialized models, including models from NVIDIA's Nemotron family through Sakana's collaboration with NVIDIA. Rather than treating the most powerful available model as the default for every request, Fugu Max dynamically routes work toward models capable of completing individual subtasks efficiently. Sakana reports Fugu Max pricing of $2 per million input tokens and $6 per million output tokens. According to the supplied release, that output pricing is 40% to 60% lower than the cited prices for Sonnet 5, GPT 5.6 Terra, and Kimi K3. The significance is not simply that one API is cheaper. It is that the cost advantage comes from an architectural strategy rather than merely reducing the price of an individual underlying model. Dimension Conventional single-model approach Orchestration approach Model selection Generally fixed Dynamically selected Task specialization Limited by the model Multiple specialized models Cost optimization Primarily model-level Workflow-level Vendor dependency Potentially concentrated Can be distributed Model replacement May require application changes Pool can be swappable Complex workflows One model handles the chain Multiple agents can collaborate This changes the economics of inference because optimization occurs across the workflow rather than solely within a model. What the Benchmark Results Suggest Fugu Max is reported to achieve the best overall score on six benchmarks, including Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish. It also reportedly expands the cost-performance Pareto frontier on seven of ten benchmarks. A Pareto frontier is useful for understanding this development because AI systems are rarely evaluated on performance alone. A model can be exceptionally capable but economically impractical, or extremely inexpensive but insufficiently capable. The desirable region is where additional spending produces meaningful additional capability. Fugu Max attempts to occupy a more efficient point within that trade-off. The broader implication is that AI competition could increasingly resemble optimization rather than a straightforward race toward a single maximum benchmark score. The winner may not always be the system with the highest raw capability. It may be the system capable of delivering a required level of capability at the lowest sustainable cost. Fugu Ultra v2 Pushes in the Opposite Direction Sakana has not treated cost efficiency as synonymous with lower capability. Fugu Ultra v2 is designed to push the other side of the frontier, maximizing performance on difficult, multi-step workloads. The supplied material reports that Fugu Ultra v2 achieves the best or joint-best result on five of eight benchmarks, including GDP.pdf, Chartography, SWEFish, DeepSWE, and Toolathon. It places in the top two on seven of the eight benchmarks. Its performance on visual and structured reasoning tasks is particularly notable. On Chartography, Fugu Ultra v2 reportedly scores 48.3, compared with 27.3 for Opus 5 and 29.5 for Fable 5. On DeepSWE, it scores 74.3, according to Sakana's published results. These figures illustrate a different proposition from Fugu Max. Fugu Max is primarily about finding an efficient location on the capability-cost frontier. Fugu Ultra v2 is about pushing the capability boundary itself. The two products therefore represent different optimization objectives built on the same underlying philosophy. The Strategic Advantage of a Swappable Model Pool One of the most important aspects of orchestration is not immediately visible in benchmark tables. It is flexibility. A single-model architecture can create dependency on one provider. If pricing changes, access is restricted, an API is modified, or a model is discontinued, applications built heavily around that model may face significant disruption. An orchestration architecture can potentially reduce that exposure by maintaining a pool of models that can be replaced or rearranged. Sakana emphasizes this characteristic as a form of supply-chain resilience. That becomes increasingly important as AI becomes critical infrastructure for businesses. Companies are beginning to use AI for software development, research, cybersecurity, customer operations, analytics, automation, and internal decision support. Dependence on one external model provider can therefore become a strategic risk rather than merely a technical preference. An orchestration layer introduces another abstraction between the application and individual models. The application asks for an outcome. The orchestration system determines how that outcome should be produced. That separation could become one of the defining architectural patterns of enterprise AI. From Model Competition to Orchestration Competition The economic implications are potentially profound. In a conventional model-centric ecosystem, providers compete to build the best model and capture value through access to it. In an orchestration-centric ecosystem, the value proposition moves upward. The orchestration provider controls how workloads are distributed across models. That gives it influence over model selection, token consumption, task decomposition, quality control, and potentially the economics of the entire workflow. This resembles the evolution seen in other technology markets where an abstraction layer eventually becomes strategically important because it controls how underlying resources are consumed. The model itself does not disappear. Instead, it becomes one component inside a larger system. This could create pressure on frontier-model providers. If users can obtain comparable results by combining several less expensive models, the premium attached to any single proprietary model could become harder to defend. That does not mean proprietary frontier models become irrelevant. The most capable systems may continue to provide critical capabilities for the hardest tasks. But their role could change from being the universal engine behind every AI workload to serving as one component in a broader computational ecosystem. Open Models Become More Valuable Through Coordination Open-weight models have traditionally faced a perception problem. Individual open models may not always match the absolute capability of the strongest proprietary systems. But orchestration changes how their value can be measured. A specialized model does not need to outperform every competitor. It only needs to be exceptionally useful for the subset of tasks for which it is selected. This produces an ecosystem in which specialization becomes an advantage. One model may be strong at coding. Another may perform better on a particular reasoning pattern. A third may offer strong visual interpretation. Another may be inexpensive enough to handle high-volume routine tasks. An intelligent orchestration layer can potentially turn these differences into a collective capability. That is strategically important because the open-model ecosystem is highly diverse and continues to expand. Instead of asking whether an open model can independently defeat the largest proprietary model, the more relevant question becomes whether a collection of specialized open models can collectively outperform a single general-purpose system on a particular workload. Fugu's Cybersecurity Direction Shows the Broader Potential Sakana's Fugu Cyber provides an example of what domain-specific orchestration can mean. The supplied material reports scores of 86.9% on CyberGym and 72.1% on CTI-REALM for the specialized cybersecurity system. The significance is architectural. Cybersecurity involves highly heterogeneous tasks, including code analysis, vulnerability reasoning, threat intelligence interpretation, system understanding, and operational decision-making. A specialized orchestration system can potentially select different capabilities depending on the problem rather than forcing every cybersecurity task through a general-purpose model. This approach could extend to other industries. Financial analysis, scientific research, legal workflows, enterprise software engineering, medical information processing, and industrial automation all contain collections of subtasks with different computational requirements. The future AI stack may therefore contain increasingly specialized orchestration systems optimized around specific domains. The Economics of AI May Become a Routing Problem The central economic question is no longer simply how much it costs to run a model. It is how much it costs to complete a useful task. Those are different measurements. Suppose an expensive model completes an entire workflow in one pass. An orchestration system could potentially divide the same workflow into multiple operations, use inexpensive models for routine stages, reserve expensive reasoning capacity for difficult stages, and use additional agents for verification. The token price of individual models then becomes only one component of the total economics. The relevant metric becomes something closer to: Cost per successful task × quality of outcome × latency × reliability That is a much more sophisticated measure of AI efficiency. It also explains why orchestration could become commercially significant even if individual underlying models continue to improve. Better models do not eliminate the need for intelligent resource allocation. In fact, greater model diversity could make orchestration more valuable because there will be more options to evaluate and select. The Remaining Challenges Orchestration is not automatically superior. Coordinating multiple models introduces complexity. Routing itself consumes computational resources. Multi-agent workflows can increase latency, generate redundant outputs, or create error propagation between agents. Evaluation also becomes harder. When a system consists of multiple models, researchers must determine whether improvements result from better underlying models, better routing, superior coordination, improved verification, or simply increased computational effort. There are also security considerations. A model-agnostic architecture can create a broader attack surface if multiple external models, APIs, tools, or agents participate in a workflow. Enterprises therefore need strong controls around data handling, model access, permissions, logging, and output validation. The orchestration layer itself becomes a critical component. If it controls which models receive sensitive information and determines how high-value tasks are processed, its reliability and governance become as important as the reliability of the individual models underneath it. The Vendor Lock-In Question Sakana's approach also raises a larger question about AI sovereignty. Vendor lock-in can emerge at multiple levels, including model APIs, proprietary tooling, data formats, agent frameworks, and application architecture. An orchestration system that supports a diverse model pool can reduce some of these dependencies. The advantage is particularly relevant for enterprises planning AI infrastructure over several years rather than several months. If a company's applications are tightly coupled to one model provider, changing providers may require significant engineering work. If applications communicate primarily with an abstraction layer, underlying models can theoretically be changed without redesigning the entire application. That creates strategic optionality. However, true portability depends on more than API compatibility. Different models have different reasoning behaviors, context handling, tool capabilities, safety characteristics, and output reliability. Switching the underlying model may still require evaluation and workflow adjustment. Orchestration reduces dependency, but it does not make dependency disappear. A New Competitive Layer in the AI Stack The release of Fugu Max and Fugu Ultra v2 points toward a future in which AI infrastructure develops several distinct layers. At the bottom sits compute infrastructure. Above it are foundation and specialized models. Above those are orchestration systems responsible for selecting, coordinating, evaluating, and combining model capabilities. At the top are applications that transform these capabilities into business outcomes. The orchestration layer could become strategically important because it connects supply with demand. It determines which model performs which task, when additional reasoning is justified, when a cheaper model is sufficient, and when multiple agents should collaborate. That makes orchestration an economic control point as much as a technical one. What This Means for Enterprises For businesses adopting AI, the emergence of orchestration changes the strategic decision-making framework. Instead of asking only which model to standardize on, enterprises may increasingly need to ask: Which workloads require frontier-level reasoning? Which tasks can be delegated to smaller models? Where does multi-agent collaboration improve reliability? How much does each successful workflow actually cost? How easily can underlying models be replaced? What data can each model access? How should outputs be evaluated and verified? This approach can make AI procurement more dynamic. A company might use several model providers simultaneously, with orchestration determining which one handles a particular workload. The result could be a more competitive AI market in which model providers compete not only on intelligence, but also on how effectively their systems participate in larger orchestrated workflows. The Future of AI May Be Collective Intelligence The most consequential idea behind Fugu is not simply that multiple models can work together. It is that intelligence itself may become increasingly compositional. The strongest AI system may not be a single model with every capability embedded inside its parameters. It may instead be a dynamic network capable of assembling the appropriate intelligence for each problem. That represents a conceptual shift from model intelligence to system intelligence. A system becomes intelligent not merely because one model is exceptionally capable, but because the overall architecture knows how to allocate computation. Fugu Max represents the cost-efficiency side of that equation. Fugu Ultra v2 represents the maximum-capability side. Together, they illustrate a market in which cost and intelligence do not necessarily have to move in opposite directions. For the broader AI industry, that could prove more disruptive than another incremental increase in benchmark performance. Conclusion Sakana AI's Fugu Max and Fugu Ultra v2 highlight a fundamental change taking place in artificial intelligence. The industry is moving from an era dominated by competition between individual foundation models toward one in which orchestration, routing, specialization, and system-level optimization become equally important. Fugu Max demonstrates the economic proposition, using a broad model pool to pursue high capability at substantially lower token costs. Fugu Ultra v2 demonstrates the performance proposition, coordinating models to target difficult reasoning, coding, visual interpretation, and agentic workloads. The strategic consequence could be significant. If orchestration becomes sufficiently reliable, enterprises may no longer need to place all of their AI workloads behind one proprietary model. They could instead operate through an intelligent layer capable of selecting among competing models according to cost, capability, domain, and availability. That would shift competitive power throughout the AI ecosystem. For technology analysts and organizations such as 1950.ai, the development represents a particularly important trend to watch. The next phase of AI may not be defined solely by who builds the largest or most capable model. It may be defined by who builds the most intelligent system for deciding which intelligence to use, when to use it, and how much it should cost. The model race is therefore unlikely to end. But increasingly, it may become only one part of a much larger race to build the infrastructure that orchestrates intelligence itself. Further Reading / External References The Orchestration Arbitrage: How Sakana’s Fugu Max Rewrites the Pricing War https://forkast.news/the-orchestration-arbitrage-how-sakanas-fugu-max-rewrites-the-pricing-war/ Introducing Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier https://sakana.ai/fugu-max-release/

The artificial intelligence industry has spent much of the past decade competing to build increasingly capable foundation models. Larger parameter counts, greater training budgets, longer context windows, and more sophisticated reasoning systems have driven successive waves of progress.


But the economics of AI are beginning to expose a different problem.

For many real-world workloads, the most capable model available is not necessarily the most economically intelligent choice. A simple classification, data lookup, code transformation, or narrowly defined reasoning task may not justify consuming the computational resources of an expensive frontier model. The emerging question is therefore shifting from which model is smartest to which combination of models can solve a particular problem most efficiently.


Sakana AI's latest Fugu releases illustrate this transition.

With Fugu Max and Fugu Ultra v2, Sakana is advancing an orchestration model in which multiple open-weight and specialized AI systems are dynamically coordinated rather than relying on one monolithic model for every task. The strategy challenges conventional assumptions about AI pricing, vendor dependence, model specialization, and where economic value will ultimately accumulate across the AI stack.

The implications extend well beyond one product release. If orchestration can consistently combine lower-cost models into systems capable of competing with more expensive frontier offerings, the competitive battlefield of AI may increasingly move from individual models to the intelligent infrastructure connecting them.


The AI Industry Is Moving Beyond the Single-Model Paradigm

The traditional foundation-model business has been relatively straightforward.

A provider develops a powerful model, operates the infrastructure required to serve it, exposes its capabilities through an API, and charges customers according to usage. Developers then build applications around that model.

This approach has obvious advantages. A single model provides a predictable interface, centralized optimization, and a relatively simple developer experience.

Its weakness is economic inefficiency.


Not every task requires the same level of intelligence. A complex software-engineering problem may demand substantial reasoning, while a straightforward extraction task may require only a smaller specialized model. Treating both workloads identically can result in unnecessary inference expenditure.

Orchestration introduces a different architecture.

Instead of asking one model to perform every operation, an orchestration layer can analyze a task, select appropriate models, divide work among agents, compare outputs, and coordinate the final result.

The fundamental architecture becomes:

Task → Routing → Specialized Models → Coordination → Verification → Final Output

That creates a new optimization target, balancing capability against cost.


Fugu Max Targets the Cost-Performance Frontier

Sakana's Fugu Max is designed around this principle.

The system orchestrates a large collection of open-weight and specialized models, including models from NVIDIA's Nemotron family through Sakana's collaboration with NVIDIA.

Rather than treating the most powerful available model as the default for every request, Fugu Max dynamically routes work toward models capable of completing individual subtasks efficiently.


Sakana reports Fugu Max pricing of $2 per million input tokens and $6 per million output tokens. According to the supplied release, that output pricing is 40% to 60% lower than the cited prices for Sonnet 5, GPT 5.6 Terra, and Kimi K3.

The significance is not simply that one API is cheaper.

It is that the cost advantage comes from an architectural strategy rather than merely reducing the price of an individual underlying model.

Dimension

Conventional single-model approach

Orchestration approach

Model selection

Generally fixed

Dynamically selected

Task specialization

Limited by the model

Multiple specialized models

Cost optimization

Primarily model-level

Workflow-level

Vendor dependency

Potentially concentrated

Can be distributed

Model replacement

May require application changes

Pool can be swappable

Complex workflows

One model handles the chain

Multiple agents can collaborate

This changes the economics of inference because optimization occurs across the workflow rather than solely within a model.


What the Benchmark Results Suggest

Fugu Max is reported to achieve the best overall score on six benchmarks, including Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish.

It also reportedly expands the cost-performance Pareto frontier on seven of ten benchmarks.

A Pareto frontier is useful for understanding this development because AI systems are rarely evaluated on performance alone. A model can be exceptionally capable but economically impractical, or extremely inexpensive but insufficiently capable.

The desirable region is where additional spending produces meaningful additional capability.


Fugu Max attempts to occupy a more efficient point within that trade-off.

The broader implication is that AI competition could increasingly resemble optimization rather than a straightforward race toward a single maximum benchmark score.

The winner may not always be the system with the highest raw capability. It may be the system capable of delivering a required level of capability at the lowest sustainable cost.


Fugu Ultra v2 Pushes in the Opposite Direction

Sakana has not treated cost efficiency as synonymous with lower capability.

Fugu Ultra v2 is designed to push the other side of the frontier, maximizing performance on difficult, multi-step workloads.

The supplied material reports that Fugu Ultra v2 achieves the best or joint-best result on five of eight benchmarks, including GDP.pdf, Chartography, SWEFish, DeepSWE, and Toolathon. It places in the top two on seven of the eight benchmarks.


Its performance on visual and structured reasoning tasks is particularly notable. On Chartography, Fugu Ultra v2 reportedly scores 48.3, compared with 27.3 for Opus 5 and 29.5 for Fable 5. On DeepSWE, it scores 74.3, according to Sakana's published results.

These figures illustrate a different proposition from Fugu Max.

Fugu Max is primarily about finding an efficient location on the capability-cost frontier. Fugu Ultra v2 is about pushing the capability boundary itself.

The two products therefore represent different optimization objectives built on the same underlying philosophy.


The Strategic Advantage of a Swappable Model Pool

One of the most important aspects of orchestration is not immediately visible in benchmark tables.

It is flexibility.

A single-model architecture can create dependency on one provider. If pricing changes, access is restricted, an API is modified, or a model is discontinued, applications built heavily around that model may face significant disruption.

An orchestration architecture can potentially reduce that exposure by maintaining a pool of models that can be replaced or rearranged.

Sakana emphasizes this characteristic as a form of supply-chain resilience.

That becomes increasingly important as AI becomes critical infrastructure for businesses.


Companies are beginning to use AI for software development, research, cybersecurity, customer operations, analytics, automation, and internal decision support. Dependence on one external model provider can therefore become a strategic risk rather than merely a technical preference.

An orchestration layer introduces another abstraction between the application and individual models.

The application asks for an outcome. The orchestration system determines how that outcome should be produced.

That separation could become one of the defining architectural patterns of enterprise AI.


From Model Competition to Orchestration Competition

The economic implications are potentially profound.

In a conventional model-centric ecosystem, providers compete to build the best model and capture value through access to it.

In an orchestration-centric ecosystem, the value proposition moves upward.

The orchestration provider controls how workloads are distributed across models. That gives it influence over model selection, token consumption, task decomposition, quality control, and potentially the economics of the entire workflow.

This resembles the evolution seen in other technology markets where an abstraction layer eventually becomes strategically important because it controls how underlying resources are consumed.


The model itself does not disappear.

Instead, it becomes one component inside a larger system.

This could create pressure on frontier-model providers. If users can obtain comparable results by combining several less expensive models, the premium attached to any single proprietary model could become harder to defend.

That does not mean proprietary frontier models become irrelevant. The most capable systems may continue to provide critical capabilities for the hardest tasks.

But their role could change from being the universal engine behind every AI workload to serving as one component in a broader computational ecosystem.


Open Models Become More Valuable Through Coordination

Open-weight models have traditionally faced a perception problem.

Individual open models may not always match the absolute capability of the strongest proprietary systems. But orchestration changes how their value can be measured.

A specialized model does not need to outperform every competitor.

It only needs to be exceptionally useful for the subset of tasks for which it is selected.

This produces an ecosystem in which specialization becomes an advantage.

One model may be strong at coding. Another may perform better on a particular reasoning pattern. A third may offer strong visual interpretation. Another may be inexpensive enough to handle high-volume routine tasks.


An intelligent orchestration layer can potentially turn these differences into a collective capability.

That is strategically important because the open-model ecosystem is highly diverse and continues to expand.

Instead of asking whether an open model can independently defeat the largest proprietary model, the more relevant question becomes whether a collection of specialized open models can collectively outperform a single general-purpose system on a particular workload.


Fugu's Cybersecurity Direction Shows the Broader Potential

Sakana's Fugu Cyber provides an example of what domain-specific orchestration can mean.

The supplied material reports scores of 86.9% on CyberGym and 72.1% on CTI-REALM for the specialized cybersecurity system.

The significance is architectural.

Cybersecurity involves highly heterogeneous tasks, including code analysis, vulnerability reasoning, threat intelligence interpretation, system understanding, and operational decision-making. A specialized orchestration system can potentially select different capabilities depending on the problem rather than forcing every cybersecurity task through a general-purpose model.


This approach could extend to other industries.

Financial analysis, scientific research, legal workflows, enterprise software engineering, medical information processing, and industrial automation all contain collections of subtasks with different computational requirements.

The future AI stack may therefore contain increasingly specialized orchestration systems optimized around specific domains.


The Economics of AI May Become a Routing Problem

The central economic question is no longer simply how much it costs to run a model.

It is how much it costs to complete a useful task.

Those are different measurements.

Suppose an expensive model completes an entire workflow in one pass. An orchestration system could potentially divide the same workflow into multiple operations, use inexpensive models for routine stages, reserve expensive reasoning capacity for difficult stages, and use additional agents for verification.

The token price of individual models then becomes only one component of the total economics.


The relevant metric becomes something closer to:

Cost per successful task × quality of outcome × latency × reliability

That is a much more sophisticated measure of AI efficiency.

It also explains why orchestration could become commercially significant even if individual underlying models continue to improve.

Better models do not eliminate the need for intelligent resource allocation. In fact, greater model diversity could make orchestration more valuable because there will be more options to evaluate and select.


The Remaining Challenges

Orchestration is not automatically superior.

Coordinating multiple models introduces complexity. Routing itself consumes computational resources. Multi-agent workflows can increase latency, generate redundant outputs, or create error propagation between agents.

Evaluation also becomes harder.


When a system consists of multiple models, researchers must determine whether improvements result from better underlying models, better routing, superior coordination, improved verification, or simply increased computational effort.

There are also security considerations.

A model-agnostic architecture can create a broader attack surface if multiple external models, APIs, tools, or agents participate in a workflow. Enterprises therefore need strong controls around data handling, model access, permissions, logging, and output validation.

The orchestration layer itself becomes a critical component.

If it controls which models receive sensitive information and determines how high-value tasks are processed, its reliability and governance become as important as the reliability of the individual models underneath it.


The artificial intelligence industry has spent much of the past decade competing to build increasingly capable foundation models. Larger parameter counts, greater training budgets, longer context windows, and more sophisticated reasoning systems have driven successive waves of progress. But the economics of AI are beginning to expose a different problem. For many real-world workloads, the most capable model available is not necessarily the most economically intelligent choice. A simple classification, data lookup, code transformation, or narrowly defined reasoning task may not justify consuming the computational resources of an expensive frontier model. The emerging question is therefore shifting from which model is smartest to which combination of models can solve a particular problem most efficiently. Sakana AI's latest Fugu releases illustrate this transition. With Fugu Max and Fugu Ultra v2, Sakana is advancing an orchestration model in which multiple open-weight and specialized AI systems are dynamically coordinated rather than relying on one monolithic model for every task. The strategy challenges conventional assumptions about AI pricing, vendor dependence, model specialization, and where economic value will ultimately accumulate across the AI stack. The implications extend well beyond one product release. If orchestration can consistently combine lower-cost models into systems capable of competing with more expensive frontier offerings, the competitive battlefield of AI may increasingly move from individual models to the intelligent infrastructure connecting them. The AI Industry Is Moving Beyond the Single-Model Paradigm The traditional foundation-model business has been relatively straightforward. A provider develops a powerful model, operates the infrastructure required to serve it, exposes its capabilities through an API, and charges customers according to usage. Developers then build applications around that model. This approach has obvious advantages. A single model provides a predictable interface, centralized optimization, and a relatively simple developer experience. Its weakness is economic inefficiency. Not every task requires the same level of intelligence. A complex software-engineering problem may demand substantial reasoning, while a straightforward extraction task may require only a smaller specialized model. Treating both workloads identically can result in unnecessary inference expenditure. Orchestration introduces a different architecture. Instead of asking one model to perform every operation, an orchestration layer can analyze a task, select appropriate models, divide work among agents, compare outputs, and coordinate the final result. The fundamental architecture becomes: Task → Routing → Specialized Models → Coordination → Verification → Final Output That creates a new optimization target, balancing capability against cost. Fugu Max Targets the Cost-Performance Frontier Sakana's Fugu Max is designed around this principle. The system orchestrates a large collection of open-weight and specialized models, including models from NVIDIA's Nemotron family through Sakana's collaboration with NVIDIA. Rather than treating the most powerful available model as the default for every request, Fugu Max dynamically routes work toward models capable of completing individual subtasks efficiently. Sakana reports Fugu Max pricing of $2 per million input tokens and $6 per million output tokens. According to the supplied release, that output pricing is 40% to 60% lower than the cited prices for Sonnet 5, GPT 5.6 Terra, and Kimi K3. The significance is not simply that one API is cheaper. It is that the cost advantage comes from an architectural strategy rather than merely reducing the price of an individual underlying model. Dimension Conventional single-model approach Orchestration approach Model selection Generally fixed Dynamically selected Task specialization Limited by the model Multiple specialized models Cost optimization Primarily model-level Workflow-level Vendor dependency Potentially concentrated Can be distributed Model replacement May require application changes Pool can be swappable Complex workflows One model handles the chain Multiple agents can collaborate This changes the economics of inference because optimization occurs across the workflow rather than solely within a model. What the Benchmark Results Suggest Fugu Max is reported to achieve the best overall score on six benchmarks, including Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish. It also reportedly expands the cost-performance Pareto frontier on seven of ten benchmarks. A Pareto frontier is useful for understanding this development because AI systems are rarely evaluated on performance alone. A model can be exceptionally capable but economically impractical, or extremely inexpensive but insufficiently capable. The desirable region is where additional spending produces meaningful additional capability. Fugu Max attempts to occupy a more efficient point within that trade-off. The broader implication is that AI competition could increasingly resemble optimization rather than a straightforward race toward a single maximum benchmark score. The winner may not always be the system with the highest raw capability. It may be the system capable of delivering a required level of capability at the lowest sustainable cost. Fugu Ultra v2 Pushes in the Opposite Direction Sakana has not treated cost efficiency as synonymous with lower capability. Fugu Ultra v2 is designed to push the other side of the frontier, maximizing performance on difficult, multi-step workloads. The supplied material reports that Fugu Ultra v2 achieves the best or joint-best result on five of eight benchmarks, including GDP.pdf, Chartography, SWEFish, DeepSWE, and Toolathon. It places in the top two on seven of the eight benchmarks. Its performance on visual and structured reasoning tasks is particularly notable. On Chartography, Fugu Ultra v2 reportedly scores 48.3, compared with 27.3 for Opus 5 and 29.5 for Fable 5. On DeepSWE, it scores 74.3, according to Sakana's published results. These figures illustrate a different proposition from Fugu Max. Fugu Max is primarily about finding an efficient location on the capability-cost frontier. Fugu Ultra v2 is about pushing the capability boundary itself. The two products therefore represent different optimization objectives built on the same underlying philosophy. The Strategic Advantage of a Swappable Model Pool One of the most important aspects of orchestration is not immediately visible in benchmark tables. It is flexibility. A single-model architecture can create dependency on one provider. If pricing changes, access is restricted, an API is modified, or a model is discontinued, applications built heavily around that model may face significant disruption. An orchestration architecture can potentially reduce that exposure by maintaining a pool of models that can be replaced or rearranged. Sakana emphasizes this characteristic as a form of supply-chain resilience. That becomes increasingly important as AI becomes critical infrastructure for businesses. Companies are beginning to use AI for software development, research, cybersecurity, customer operations, analytics, automation, and internal decision support. Dependence on one external model provider can therefore become a strategic risk rather than merely a technical preference. An orchestration layer introduces another abstraction between the application and individual models. The application asks for an outcome. The orchestration system determines how that outcome should be produced. That separation could become one of the defining architectural patterns of enterprise AI. From Model Competition to Orchestration Competition The economic implications are potentially profound. In a conventional model-centric ecosystem, providers compete to build the best model and capture value through access to it. In an orchestration-centric ecosystem, the value proposition moves upward. The orchestration provider controls how workloads are distributed across models. That gives it influence over model selection, token consumption, task decomposition, quality control, and potentially the economics of the entire workflow. This resembles the evolution seen in other technology markets where an abstraction layer eventually becomes strategically important because it controls how underlying resources are consumed. The model itself does not disappear. Instead, it becomes one component inside a larger system. This could create pressure on frontier-model providers. If users can obtain comparable results by combining several less expensive models, the premium attached to any single proprietary model could become harder to defend. That does not mean proprietary frontier models become irrelevant. The most capable systems may continue to provide critical capabilities for the hardest tasks. But their role could change from being the universal engine behind every AI workload to serving as one component in a broader computational ecosystem. Open Models Become More Valuable Through Coordination Open-weight models have traditionally faced a perception problem. Individual open models may not always match the absolute capability of the strongest proprietary systems. But orchestration changes how their value can be measured. A specialized model does not need to outperform every competitor. It only needs to be exceptionally useful for the subset of tasks for which it is selected. This produces an ecosystem in which specialization becomes an advantage. One model may be strong at coding. Another may perform better on a particular reasoning pattern. A third may offer strong visual interpretation. Another may be inexpensive enough to handle high-volume routine tasks. An intelligent orchestration layer can potentially turn these differences into a collective capability. That is strategically important because the open-model ecosystem is highly diverse and continues to expand. Instead of asking whether an open model can independently defeat the largest proprietary model, the more relevant question becomes whether a collection of specialized open models can collectively outperform a single general-purpose system on a particular workload. Fugu's Cybersecurity Direction Shows the Broader Potential Sakana's Fugu Cyber provides an example of what domain-specific orchestration can mean. The supplied material reports scores of 86.9% on CyberGym and 72.1% on CTI-REALM for the specialized cybersecurity system. The significance is architectural. Cybersecurity involves highly heterogeneous tasks, including code analysis, vulnerability reasoning, threat intelligence interpretation, system understanding, and operational decision-making. A specialized orchestration system can potentially select different capabilities depending on the problem rather than forcing every cybersecurity task through a general-purpose model. This approach could extend to other industries. Financial analysis, scientific research, legal workflows, enterprise software engineering, medical information processing, and industrial automation all contain collections of subtasks with different computational requirements. The future AI stack may therefore contain increasingly specialized orchestration systems optimized around specific domains. The Economics of AI May Become a Routing Problem The central economic question is no longer simply how much it costs to run a model. It is how much it costs to complete a useful task. Those are different measurements. Suppose an expensive model completes an entire workflow in one pass. An orchestration system could potentially divide the same workflow into multiple operations, use inexpensive models for routine stages, reserve expensive reasoning capacity for difficult stages, and use additional agents for verification. The token price of individual models then becomes only one component of the total economics. The relevant metric becomes something closer to: Cost per successful task × quality of outcome × latency × reliability That is a much more sophisticated measure of AI efficiency. It also explains why orchestration could become commercially significant even if individual underlying models continue to improve. Better models do not eliminate the need for intelligent resource allocation. In fact, greater model diversity could make orchestration more valuable because there will be more options to evaluate and select. The Remaining Challenges Orchestration is not automatically superior. Coordinating multiple models introduces complexity. Routing itself consumes computational resources. Multi-agent workflows can increase latency, generate redundant outputs, or create error propagation between agents. Evaluation also becomes harder. When a system consists of multiple models, researchers must determine whether improvements result from better underlying models, better routing, superior coordination, improved verification, or simply increased computational effort. There are also security considerations. A model-agnostic architecture can create a broader attack surface if multiple external models, APIs, tools, or agents participate in a workflow. Enterprises therefore need strong controls around data handling, model access, permissions, logging, and output validation. The orchestration layer itself becomes a critical component. If it controls which models receive sensitive information and determines how high-value tasks are processed, its reliability and governance become as important as the reliability of the individual models underneath it. The Vendor Lock-In Question Sakana's approach also raises a larger question about AI sovereignty. Vendor lock-in can emerge at multiple levels, including model APIs, proprietary tooling, data formats, agent frameworks, and application architecture. An orchestration system that supports a diverse model pool can reduce some of these dependencies. The advantage is particularly relevant for enterprises planning AI infrastructure over several years rather than several months. If a company's applications are tightly coupled to one model provider, changing providers may require significant engineering work. If applications communicate primarily with an abstraction layer, underlying models can theoretically be changed without redesigning the entire application. That creates strategic optionality. However, true portability depends on more than API compatibility. Different models have different reasoning behaviors, context handling, tool capabilities, safety characteristics, and output reliability. Switching the underlying model may still require evaluation and workflow adjustment. Orchestration reduces dependency, but it does not make dependency disappear. A New Competitive Layer in the AI Stack The release of Fugu Max and Fugu Ultra v2 points toward a future in which AI infrastructure develops several distinct layers. At the bottom sits compute infrastructure. Above it are foundation and specialized models. Above those are orchestration systems responsible for selecting, coordinating, evaluating, and combining model capabilities. At the top are applications that transform these capabilities into business outcomes. The orchestration layer could become strategically important because it connects supply with demand. It determines which model performs which task, when additional reasoning is justified, when a cheaper model is sufficient, and when multiple agents should collaborate. That makes orchestration an economic control point as much as a technical one. What This Means for Enterprises For businesses adopting AI, the emergence of orchestration changes the strategic decision-making framework. Instead of asking only which model to standardize on, enterprises may increasingly need to ask: Which workloads require frontier-level reasoning? Which tasks can be delegated to smaller models? Where does multi-agent collaboration improve reliability? How much does each successful workflow actually cost? How easily can underlying models be replaced? What data can each model access? How should outputs be evaluated and verified? This approach can make AI procurement more dynamic. A company might use several model providers simultaneously, with orchestration determining which one handles a particular workload. The result could be a more competitive AI market in which model providers compete not only on intelligence, but also on how effectively their systems participate in larger orchestrated workflows. The Future of AI May Be Collective Intelligence The most consequential idea behind Fugu is not simply that multiple models can work together. It is that intelligence itself may become increasingly compositional. The strongest AI system may not be a single model with every capability embedded inside its parameters. It may instead be a dynamic network capable of assembling the appropriate intelligence for each problem. That represents a conceptual shift from model intelligence to system intelligence. A system becomes intelligent not merely because one model is exceptionally capable, but because the overall architecture knows how to allocate computation. Fugu Max represents the cost-efficiency side of that equation. Fugu Ultra v2 represents the maximum-capability side. Together, they illustrate a market in which cost and intelligence do not necessarily have to move in opposite directions. For the broader AI industry, that could prove more disruptive than another incremental increase in benchmark performance. Conclusion Sakana AI's Fugu Max and Fugu Ultra v2 highlight a fundamental change taking place in artificial intelligence. The industry is moving from an era dominated by competition between individual foundation models toward one in which orchestration, routing, specialization, and system-level optimization become equally important. Fugu Max demonstrates the economic proposition, using a broad model pool to pursue high capability at substantially lower token costs. Fugu Ultra v2 demonstrates the performance proposition, coordinating models to target difficult reasoning, coding, visual interpretation, and agentic workloads. The strategic consequence could be significant. If orchestration becomes sufficiently reliable, enterprises may no longer need to place all of their AI workloads behind one proprietary model. They could instead operate through an intelligent layer capable of selecting among competing models according to cost, capability, domain, and availability. That would shift competitive power throughout the AI ecosystem. For technology analysts and organizations such as 1950.ai, the development represents a particularly important trend to watch. The next phase of AI may not be defined solely by who builds the largest or most capable model. It may be defined by who builds the most intelligent system for deciding which intelligence to use, when to use it, and how much it should cost. The model race is therefore unlikely to end. But increasingly, it may become only one part of a much larger race to build the infrastructure that orchestrates intelligence itself. Further Reading / External References The Orchestration Arbitrage: How Sakana’s Fugu Max Rewrites the Pricing War https://forkast.news/the-orchestration-arbitrage-how-sakanas-fugu-max-rewrites-the-pricing-war/ Introducing Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier https://sakana.ai/fugu-max-release/

The Vendor Lock-In Question

Sakana's approach also raises a larger question about AI sovereignty.

Vendor lock-in can emerge at multiple levels, including model APIs, proprietary tooling, data formats, agent frameworks, and application architecture.

An orchestration system that supports a diverse model pool can reduce some of these dependencies.


The advantage is particularly relevant for enterprises planning AI infrastructure over several years rather than several months.

If a company's applications are tightly coupled to one model provider, changing providers may require significant engineering work. If applications communicate primarily with an abstraction layer, underlying models can theoretically be changed without redesigning the entire application.

That creates strategic optionality.

However, true portability depends on more than API compatibility. Different models have different reasoning behaviors, context handling, tool capabilities, safety characteristics, and output reliability. Switching the underlying model may still require evaluation and workflow adjustment.

Orchestration reduces dependency, but it does not make dependency disappear.


A New Competitive Layer in the AI Stack

The release of Fugu Max and Fugu Ultra v2 points toward a future in which AI infrastructure develops several distinct layers.

At the bottom sits compute infrastructure.

Above it are foundation and specialized models.

Above those are orchestration systems responsible for selecting, coordinating, evaluating, and combining model capabilities.


At the top are applications that transform these capabilities into business outcomes.

The orchestration layer could become strategically important because it connects supply with demand.

It determines which model performs which task, when additional reasoning is justified, when a cheaper model is sufficient, and when multiple agents should collaborate.

That makes orchestration an economic control point as much as a technical one.


What This Means for Enterprises

For businesses adopting AI, the emergence of orchestration changes the strategic decision-making framework.

Instead of asking only which model to standardize on, enterprises may increasingly need to ask:

  • Which workloads require frontier-level reasoning?

  • Which tasks can be delegated to smaller models?

  • Where does multi-agent collaboration improve reliability?

  • How much does each successful workflow actually cost?

  • How easily can underlying models be replaced?

  • What data can each model access?

  • How should outputs be evaluated and verified?

This approach can make AI procurement more dynamic.

A company might use several model providers simultaneously, with orchestration determining which one handles a particular workload.

The result could be a more competitive AI market in which model providers compete not only on intelligence, but also on how effectively their systems participate in larger orchestrated workflows.


The Future of AI May Be Collective Intelligence

The most consequential idea behind Fugu is not simply that multiple models can work together.

It is that intelligence itself may become increasingly compositional.

The strongest AI system may not be a single model with every capability embedded inside its parameters. It may instead be a dynamic network capable of assembling the appropriate intelligence for each problem.


That represents a conceptual shift from model intelligence to system intelligence.

A system becomes intelligent not merely because one model is exceptionally capable, but because the overall architecture knows how to allocate computation.

Fugu Max represents the cost-efficiency side of that equation. Fugu Ultra v2 represents the maximum-capability side. Together, they illustrate a market in which cost and intelligence do not necessarily have to move in opposite directions.

For the broader AI industry, that could prove more disruptive than another incremental increase in benchmark performance.


Conclusion

Sakana AI's Fugu Max and Fugu Ultra v2 highlight a fundamental change taking place in artificial intelligence.

The industry is moving from an era dominated by competition between individual foundation models toward one in which orchestration, routing, specialization, and system-level optimization become equally important.


Fugu Max demonstrates the economic proposition, using a broad model pool to pursue high capability at substantially lower token costs. Fugu Ultra v2 demonstrates the performance proposition, coordinating models to target difficult reasoning, coding, visual interpretation, and agentic workloads.

The strategic consequence could be significant.

If orchestration becomes sufficiently reliable, enterprises may no longer need to place all of their AI workloads behind one proprietary model. They could instead operate through an intelligent layer capable of selecting among competing models according to cost, capability, domain, and availability.


That would shift competitive power throughout the AI ecosystem.

For technology analysts and organizations such as 1950.ai, the development represents a particularly important trend to watch. The next phase of AI may not be defined solely by who builds the largest or most capable model. It may be defined by who builds the most intelligent system for deciding which intelligence to use, when to use it, and how much it should cost.

The model race is therefore unlikely to end. But increasingly, it may become only one part of a much larger race to build the infrastructure that orchestrates intelligence itself.


Further Reading / External References

The Orchestration Arbitrage: How Sakana’s Fugu Max Rewrites the Pricing War

Introducing Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier

Comments

bottom of page