Sakana AI Splits Fugu Into Max and Ultra v2 to Cut Costs 60%
Sakana AI's Fugu Max and Fugu Ultra v2 orchestrate pools of open models to beat frontier LLMs on cost and capability simultaneously.
- Sakana AI released Fugu Max and Fugu Ultra v2, two orchestration variants of its multi-agent system.
- Fugu Max prices at $2/$6 per million input/output tokens, 40-60% below Sonnet 5, GPT 5.6 Terra, Kimi K3.
- Fugu Max expands the Pareto frontier on seven of ten benchmarks including Terminal Bench 2.1 and GPQAD.
- Fugu Ultra v2 scores 48.3 on Chartography vs Opus 5's 27.3, and 74.3 on DeepSWE.
- Ultra v2 hits these numbers without Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool.
- Available today via OpenAI-compatible API at sakana.ai/fugu, one-line upgrade for existing users.
Sakana AI splits Fugu between lower costs and harder tasks
Sakana AI has released Fugu Max and Fugu Ultra v2, two configurations of its multi-agent orchestration architecture. Max targets lower-cost output, while Ultra v2 targets peak performance on complex, multi-step work. Both dynamically route requests across a swappable pool of open-weight and specialized models.
The architecture moves model selection from application code into a routing layer. Developers send requests through one API, and Fugu selects the model it expects can complete each task with the desired balance of cost and quality.
All pricing and benchmark comparisons below come from Sakana’s release announcement and remain vendor-reported until independently reproduced.
One router, two optimization targets
Sakana evaluates the variants along a Pareto frontier, the set of options where improving quality generally requires more spending and reducing cost generally sacrifices quality. Max tries to move that boundary by assigning simpler work to cheaper models. Ultra v2 spends more routing and inference capacity on tasks that demand stronger reasoning, tool use, coding, or visual analysis.
| Variant | Primary target | Workload profile |
|---|---|---|
| Fugu Max | Lower cost at competitive benchmark quality | High-volume pipelines with mixed task difficulty |
| Fugu Ultra v2 | Highest available quality from the model pool | Long, complex tasks involving reasoning, tools, code, or charts |
Max routes work to cheaper models
Fugu Max draws from a broad pool of open-weight and specialized models, including NVIDIA’s Nemotron family through a collaboration with NVIDIA. Its router assigns each task to the smallest model expected to solve it, reducing the use of more expensive models on routine requests.
Sakana lists Max at $2 per million input tokens and $6 per million output tokens. The company says its output rate is 40% to 60% lower than Sonnet 5, GPT 5.6 Terra, and Kimi K3.
- Max reportedly led six benchmarks: Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish, Sakana’s internal coding benchmark.
- It extended the reported cost-performance Pareto frontier on seven of ten benchmarks, meaning no tested competitor matched its combination of score and price on those evaluations.
- Its strongest fit is high-volume work such as extraction, data lookup, routine coding, and pipelines whose requests vary widely in difficulty.
Ultra v2 raises the benchmark ceiling
Fugu Ultra v2 is tuned for long, irregular tasks where routing for quality takes priority over minimizing token spend. Sakana says its agent pool excluded Fable 5, Fable 5.1, and GPT-6-Astra during the reported evaluations, so those proprietary models did not contribute to its scores.
On Chartography, a visual-reasoning and data-interpretation benchmark, Ultra v2 scored 48.3. Sakana reports scores of 27.3 for Opus 5 and 29.5 for Fable 5 on the same test. On DeepSWE, which evaluates software-engineering work, Ultra v2 scored 74.3 and exceeded compared models that cost three to five times more per token.
Across the eight-benchmark suite, Ultra v2 recorded the best or joint-best result on GDP.pdf, Chartography, SWEFish, DeepSWE, and Toolathon. It placed among the top two systems on seven of the eight tests.
A model pool spreads provider risk
A swappable pool gives Sakana room to replace models when providers change prices, deprecate APIs, restrict regions, or experience capacity problems. Routing changes can absorb some of those events without forcing developers to rewrite the application layer.
Operational resilience still depends on the quality of fallback routes, available capacity, provider diversity, and how quickly Sakana updates the pool. Model substitutions can also change latency, output style, tool behavior, and reproducibility, even when the API contract remains stable.
One API, two deployment profiles
Both variants are available through an OpenAI-compatible API. Sakana says existing Fugu users can upgrade with a one-line parameter change while retaining the same API contract. Access is available through the Fugu page and Sakana’s console.
- Fugu Max fits production systems where aggregate token spend matters and many requests can be handled by smaller models.
- Fugu Ultra v2 fits autonomous research, full-stack software development, multi-step tool use, and visual or chart-heavy analysis.
Checks before production
Production evaluation should cover behavior that aggregate benchmark scores and token prices cannot establish on their own.
| Area | What to verify |
|---|---|
| Latency | Median and tail latency by workload, including multi-step routes |
| Routing controls | Whether teams can pin models, inspect routes, or receive notice when the pool changes |
| Compatibility | Tool calling, structured outputs, streaming, context limits, and error semantics |
| Data governance | Retention, residency, subprocessors, model-provider access, and training policies |
| Reliability | Rate limits, service-level commitments, fallback behavior, and regional availability |
| Evaluation | Prompts, scoring methods, run variance, token use, and reproducible benchmark settings |
The router becomes a stack layer
Sakana’s recent release sequence shows how the company has expanded Fugu from an orchestration experiment into a broader product line.
- April: Fugu entered beta as a unified interface for multi-agent orchestration.
- June: Sakana moved Fugu to general availability and released Ultra v1.
- July: The company added Fugu-Cyber and a Claude Code interface.
- August: Sakana launched Sakana Chat and integrated NVIDIA Nemotron models.
- Current release: Max adds a cost-focused profile, while Ultra v2 extends the quality-focused profile.
Fugu’s product-level choice now concerns the routing policy, while the system chooses an underlying model for each task. Its production advantage will depend on latency, consistency, operational controls, and whether independent testing reproduces Sakana’s benchmark and cost claims.
Comments
