Full text · 8,024 chars
- Sakana AI launched Fugu Ultra v2 on OpenRouter, a learned orchestration engine over a swappable model pool.
- Priced at $5/$30 per million input/output tokens with a 1M-token context window and August 2026 knowledge cutoff.
- Best or joint-best on 5 of 8 benchmarks including DeepSWE (74.3), Chartography (48.3), SWEFish, GDP.pdf, and Toolathon.
- Chartography score of 48.3 beats Opus 5 (27.3) and Fable 5 (29.5) on visual and structured-data reasoning.
- Achieves these scores without Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool, per Sakana's release.
- OpenAI-compatible API, supports tool calling, structured outputs, PDF/image input, web search, and configurable reasoning effort.
Sakana Brings Fugu Ultra v2’s Model Router to OpenRouter
Sakana AI has released Fugu Ultra v2 on OpenRouter, giving developers access to a learned orchestration system that assigns each subtask to a model in a configured agent pool. The quality-focused system targets multi-step reasoning, autonomous research, and full-stack software development.
Sakana says the pool excludes Fable 5, Fable 5.1, and GPT-6-Astra. Its reported benchmark results therefore do not depend on those closed frontier models. The architecture gives Sakana room to replace individual agents as models, prices, and availability change.
Inside Fugu’s Recursive Router
Fugu uses a controller language model to analyze a request, break it into subtasks, and select a specialist for each one. The controller can also invoke additional instances of itself for planning and decomposition, creating a recursive workflow behind a single API request.
The agent pool is fixed for a given system configuration, while its members remain replaceable between configurations. Applications interact with one endpoint, and Sakana manages the internal routing. This design can reduce dependence on any individual model API, although customers still rely on Sakana to operate the hosted service.
Vendor Results Favor Multi-Step Tasks
Sakana presents Fugu Max and Fugu Ultra v2 as two configurations of the same architecture. Fugu Max optimizes output quality relative to cost. Ultra v2 prioritizes the highest available capability on complex tasks.
According to Sakana’s published results, Ultra v2 performs best on evaluations that combine long-horizon reasoning with software, visual, or structured data. These are vendor-reported scores and should be independently reproduced before they guide production decisions.
| Reported Fugu Ultra v2 benchmark results | | |
|---|---|---|
| Measure | Reported result | Context |
|---|---|---|
| Overall coverage | Best or joint-best on five of eight benchmarks | GDP.pdf, Chartography, SWEFish, DeepSWE, and Toolathon |
| Top-two coverage | Seven of eight benchmarks | Measures consistency across tasks involving planning, tools, and multiple steps |
| Chartography | 48.3 | Opus 5 scored 27.3; Fable 5 scored 29.5 |
| DeepSWE | 74.3 | Sakana says competing models cost three to five times more per token |
Chartography produced the largest reported gap, with Ultra v2 leading Opus 5 by 21 points and Fable 5 by 18.8 points on the benchmark’s scale. The evaluation tests visual reasoning and data interpretation. DeepSWE measures performance on real-world software engineering work.
The Meter Includes Internal Work
Fugu Ultra v2 is available through OpenRouter with Sakana as its sole provider. The listing specifies the following limits and rates:
| Fugu Ultra v2 access and pricing | |
|---|---|
| Input | $5 per million tokens |
|---|---|
| Output | $30 per million tokens |
| Context window | 1 million tokens |
| Long prompts | Prompts above 272,000 tokens use a higher rate |
| Knowledge cutoff | August 2026 |
| Listed release | September 11 |
| Provider | Sakana AI |
Internal orchestration tokens are billed as standard input and output tokens. A multi-step request can therefore consume more billable tokens than a direct, single-model call. Teams should compare systems using total task cost, completion rate, and latency instead of headline token prices.
Built-in web search can retrieve information newer than the listed knowledge cutoff. The cutoff describes the model’s embedded knowledge, while search freshness depends on the sources available during a request.
A One-Line Model Swap
Fugu Ultra v2 and Fugu Max use an OpenAI-compatible API. Applications already sending OpenAI SDK requests through OpenRouter can make a basic migration by changing the model slug:
from openai import OpenAI
client = OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="..."
)
response = client.chat.completions.create(
model="sakana/fugu-ultra-v2",
messages=[
{"role": "user", "content": "Refactor this repo..."}
],
)
Protocol compatibility simplifies the initial integration, but production migrations still require regression tests for tool behavior, structured output, latency, token consumption, and response quality.
Controls for Tools and Reasoning
Fugu Ultra v2 exposes features aimed at agent workflows:
- Reasoning effort: The supported levels are high ,xhigh , andmax . The setting controls how aggressively the orchestrator plans and invokes sub-agents.
- Function calling: Applications can connect the model to external tools and APIs.
- Structured outputs: Responses can follow machine-readable formats for downstream processing.
- Document input: The model accepts images and PDFs for visual and structured-data tasks.
- Web search: The system can retrieve current external information during a run.
Reasoning effort is the main control for balancing quality, token use, and latency. Developers should test all three levels on representative tasks because additional planning and delegation can raise costs without improving every request.
Tasks That Benefit From Routing
Long-running workflows can benefit when their steps require different capabilities. Sakana highlights three main categories:
- Autonomous software work: Separate agents can plan changes, inspect a repository, write code, run tools, and review results.
- Deep research: The controller can divide a question into searches, source analysis, synthesis, and verification.
- Document-heavy analysis: Specialist models can handle PDFs, charts, images, and structured data within one workflow.
Selective routing can also keep straightforward subtasks on leaner models. The resulting savings depend on the controller’s choices and the amount of internal communication required to finish the job.
Questions to Settle Before Production
Production evaluation should account for the system’s orchestration layer as well as its final answers. Teams considering Fugu Ultra v2 should:
- Measure total billed tokens and wall-clock latency for complete tasks.
- Compare reasoning-effort settings against the same acceptance criteria.
- Test tool failures, malformed structured output, and partial task completion.
- Confirm the higher rate for prompts above 272,000 tokens.
- Review rate limits, routing visibility, data retention, and retry behavior with Sakana.
- Run independent evaluations that match the application’s repositories, documents, and tools.
OpenRouter currently lists only Sakana as the provider, so the endpoint lacks provider-level redundancy. The replaceable internal agent pool may reduce model-specific exposure, while service availability, billing, and routing remain under Sakana’s control.
Sakana’s Portfolio Thesis
Sakana describes model selection as a Pareto surface, a set of options that trade cost against capability. Its thesis is that a learned router can choose among specialists more efficiently than an application that sends every task through the same model.
Fugu Ultra v2 supplies vendor evidence for that approach, particularly on software, chart, and document benchmarks. Independent testing will determine whether those gains persist across production workloads, where latency, tool reliability, token overhead, and failure recovery matter alongside benchmark scores.