01
Overview
Sakana AI released Fugu Max v1.0 and Fugu Ultra v2.0 on September 11, 2026, expanding its Fugu product from a single orchestration concept into clearer cost- and capability-focused options. Unlike a conventional foundation model that answers with one fixed model, Sakana Fugu is an orchestration system: it coordinates a pool of expert models and decides how to route and combine work for each request.
02
What launched?
Fugu Max is the cost-performance-focused option. Sakana says it expands the available pool of open-weight and specialist models, including NVIDIA Nemotron models, and dynamically selects combinations intended to solve a task efficiently. Fugu Ultra v2 is the higher-capability option for difficult multi-step reasoning, autonomous research, visual/structured-data analysis and software engineering.
Both are available through Sakana's standard OpenAI-compatible API. Existing Fugu users can switch by changing the model parameter rather than rewriting their integration.
03
API model IDs
Sakana's current pricing documentation lists fugu-max-v1.0 for Fugu Max and fugu-ultra-v2.0 for Fugu Ultra v2. Because the API is OpenAI-compatible, developers can point an existing compatible client or coding harness at Sakana's endpoint and select the desired model.
That compatibility reduces migration work, but it does not mean behavior is identical to an OpenAI model. Fugu may invoke multiple agents internally, so latency, cost structure and output characteristics can differ substantially from a single-model request.
04
Fugu Max pricing
Sakana currently prices Fugu Max at $2 per million input tokens, $6 per million output tokens and $0.25 per million cached-input tokens. The price is fixed regardless of context length according to the current pricing page. Built-in web_search and web_fetch calls are billed separately at $0.007 per call.
These are first-party prices and can change, so production budgeting should always recheck Sakana's live pricing page.
05
Fugu Ultra v2 pricing
Fugu Ultra v2 costs $5 per million input tokens, $30 per million output tokens and $0.50 per million cached-input tokens at standard context lengths. For context above 272K tokens, Sakana lists $10 per million input tokens, $45 per million output tokens and $1 per million cached-input token.
The large jump in output cost reflects the product's positioning: Ultra v2 is intended for harder workloads where quality is prioritized over minimum token spend.
06
How Sakana Fugu works
Sakana Fugu is based on model orchestration rather than a monolithic model. Sakana says its approach builds on its TRINITY and Conductor research, where a coordinator learns how to assemble, route and coordinate expert agents instead of relying on a manually designed fixed workflow.
The practical implication is that a request can be decomposed and delegated to different models. The system then combines the work into one result. Fugu Max emphasizes efficient combinations of models; Fugu Ultra uses a deeper fixed pool to maximize answer quality.
Sakana also says users of the standard Fugu product can opt specific models out of the pool for privacy and compliance reasons, while Fugu Max and Fugu Ultra use fixed pools. Enterprises needing custom provider restrictions are directed to contact Sakana.
07
Fugu Max vs Fugu Ultra v2
Choose Fugu Max when cost-performance efficiency matters most and you still want orchestration across specialist models. Choose Fugu Ultra v2 when the task is difficult enough that higher output quality is worth a substantially higher token price.
That makes Max more natural for recurring coding, analysis, automation and agent workloads where usage can scale quickly. Ultra v2 is better positioned for complex research, difficult engineering tasks, long multi-step reasoning and high-value work where a weaker answer is more expensive than additional inference cost.
Neither option should be treated as universally superior. The right choice depends on the workload, acceptable latency, model-provider constraints, and how much value additional orchestration creates for that specific task.
08
Benchmarks: what Sakana claims
Sakana reports that Fugu Max achieved the best overall score in its comparison on six benchmarks, including Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench and SWEFish. It also says Max expands the cost-performance Pareto frontier on seven of ten evaluated benchmarks.
For Fugu Ultra v2, Sakana highlights a 48.3 score on Chartography and 74.3 on DeepSWE, along with strong results on its SWEFish benchmark.
These results are useful evidence for how Sakana positions the products, but they are vendor-run evaluations. They should not be treated as independent proof that Fugu Max or Ultra v2 is universally better than Claude, GPT, Gemini, Kimi or another system. Real-world evaluation on representative workloads remains necessary.
09
Why the release matters
The interesting part of Fugu is not simply another model name. It is a different abstraction for buying AI capability. Instead of selecting one model and building orchestration yourself, developers can call one API model that decides which expert systems to involve internally.
If that approach works reliably, it creates a middle layer between model providers and applications: applications buy an outcome-oriented orchestration service while the provider changes the underlying model pool over time. That could reduce single-vendor dependency, but it also means developers give the orchestrator more control over which systems handle a request.
10
Availability
Fugu Max and Fugu Ultra v2 are available now through Sakana's OpenAI-compatible API. Sakana also offers monthly subscription plans for lighter or interactive usage, while its token-based pay-as-you-go pricing is positioned for heavier production workloads.
The current product page lists Standard at $20/month, Pro at $100/month and Max at $200/month. Those subscription tiers are usage allowances for the Fugu service, not the same thing as the Fugu Max model name. The naming can be confusing, so users should distinguish the $200/month Max subscription tier from fugu-max-v1.0, the orchestration model billed by tokens in pay-as-you-go usage.
11
Important limitations
The exact model pool can evolve. Sakana publicly names some families, but developers should not assume every request uses a particular underlying model unless Sakana documents it. Fugu Max and Ultra use fixed pools, which may matter for regulated workloads. Sakana does not publish one universal latency figure, request-rate limit or guaranteed cost per completed task because internal orchestration can vary by prompt.
The benchmark results are first-party. Cost comparisons against competitor models also depend on the workload, number of generated tokens, tool usage and whether orchestration improves first-pass success enough to offset extra internal work.
12
FAQ
Is Sakana Fugu Max a single LLM?
No. Sakana presents Fugu as a multi-agent orchestration system exposed as a model-like API. It coordinates other expert models behind one endpoint.
What is the Fugu Max model ID?
The current pricing page lists fugu-max-v1.0.
What is the Fugu Ultra v2 model ID?
The current pricing page lists fugu-ultra-v2.0.
Is the API OpenAI-compatible?
Yes. Sakana says Fugu is available through an OpenAI-compatible API so existing compatible clients and coding harnesses can migrate with minimal integration changes.
How much does Fugu Max cost?
Current pay-as-you-go pricing is $2/M input tokens, $6/M output tokens and $0.25/M cached-input tokens. Web search and fetch tools cost extra.
How much does Fugu Ultra v2 cost?
Current standard pricing is $5/M input, $30/M output and $0.50/M cached input. Rates increase above 272K context.
Which should I use?
Fugu Max is the lower-cost, efficiency-focused choice; Fugu Ultra v2 prioritizes maximum quality on difficult work. Benchmark both on your own tasks before standardizing.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.