Sakana AI's Fugu Ultra Beats Frontier Models by Coordinating Them
Sakana AI's Fugu Ultra is now on OpenRouter — a 7B orchestrator that routes across frontier models to match Claude Fable 5 and Mythos on benchmarks, without training a single frontier model.

- Fugu Ultra on OpenRouter: Sakana AI's multi-agent orchestration model is now accessible via OpenRouter, using an OpenAI-compatible API.
- 7B orchestrator beats frontier models: A 7B parameter model routes tasks across a pool of larger frontier models, matching or beating Claude Fable 5 and Mythos Preview on key benchmarks.
- Top benchmark scores: Fugu Ultra hits 93.2 on LiveCodeBench, 95.1 on GPQA-Diamond, and 54.2 on SWE-Pro — competitive with GPT-5.4 and Gemini 3.1.
- Pricing: $5/M input tokens, $30/M output tokens; subscription plans from $20–$200/month. Not yet available in EU/EEA.
- Geopolitical hedge: Architecture is designed to swap out any restricted provider dynamically, positioning it as an AI sovereignty play for non-US enterprises.
- Weaknesses remain: Lags behind GPT-5.5 on long-context recall and Opus 4.8 on cybersecurity benchmarks; output can contain logic errors on complex creative tasks.
Sakana AI just made Fugu Ultra available on OpenRouter, extending access to what may be the most interesting architectural bet in AI right now: a relatively small model whose entire job is to coordinate other, larger models. The result, according to benchmarks, is frontier-level performance , without ever training a frontier model.
One API, a whole team behind it
Sakana Fugu is a new product that delivers a full multi-agent orchestration system as a single foundation model. Fugu dynamically orchestrates the world's best models to tackle complex, multi-step tasks, accessible through a single model API. From your code's perspective, you're talking to one endpoint. Behind the scenes, a coordinated team of specialist models is doing the actual work.
Instead of training a new frontier model from scratch, Sakana trained a 7-billion-parameter orchestrator whose job is to decide which external model should handle each part of a problem. It can call a pool of third-party systems, delegate sub-tasks, check outputs, and synthesize one answer through an OpenAI-compatible endpoint.
Sakana Fugu is itself a language model trained to call various LLMs in an agent pool, including instances of itself recursively. That recursive self-calling is key: the orchestrator can spawn sub-orchestrators for deeply nested tasks, not just hand off to a flat list of specialists.
Two tiers, one integration
Sakana Fugu comes in two models , Fugu and Fugu Ultra , both available through one OpenAI-compatible API. Pick the model that fits your workload, or switch between them without changing your integration.
- Fugu: Balances strong performance with low latency, making it a great default for everyday work. It fits naturally into tools like Codex for coding and code review, as well as chatbots and other interactive services. For teams with data, privacy, or compliance requirements, Fugu also lets you opt specific agents out of its pool.
- Fugu Ultra: Tuned for maximum answer quality on hard, multi-step problems, coordinating a deeper pool of expert agents when accuracy and depth matter most. Early users have relied on it for demanding work such as AI research, paper reproduction, cybersecurity analysis, and literature and patent investigations.
Benchmark numbers that raise eyebrows
Fugu Ultra scored 54.2 on SWE-Pro, 95.1 on GPQA-Diamond, and 93.2 on LiveCodeBench v6, putting it around the same range as Claude Opus 4.6, Gemini 3.1, and GPT-5.4 on those tests. For context, GPQA-Diamond is a set of 198 graduate-level multiple-choice questions in biology, physics, and chemistry , a notoriously hard benchmark for reasoning.
Fugu actually exceeds the performance of Anthropic's Claude Fable 5 on LiveCodeBench (Fugu Ultra: 93.2, Fugu: 92.9, Fable: 89.8), and beats the prior Claude Mythos Preview model on GPQA-Diamond (Fugu Ultra: 95.5, Fugu: 95.5, Mythos Preview: 94.6).
There are limits, though. On the MRCRv2 long-context-recall test, OpenAI's GPT-5.5 maintained the lead (94.8 vs Fugu Ultra's 93.6), and Opus 4.8 remained the top performer on the CTI-REALM cybersecurity benchmark (69.6 vs Fugu Ultra's 69.4). Fugu is highly effective at boosting performance on messy, multi-step tasks by leaning on the combined strengths of multiple models. However, for sheer brute-force reasoning within a single, highly constrained domain, the industry's largest standalone models still hold the edge.
What it's actually built on
The architecture isn't improvised , it's grounded in two ICLR 2026 papers from Sakana's research team: TRINITY, which covers evolved LLM coordination, and Conductor, which focuses on learning to orchestrate agents in natural language. The key insight from both: orchestration behavior should be learned, not hard-coded. Rather than writing rules for when to delegate, the orchestrator model figures it out from training.
Because Fugu is built on learned orchestration rather than fixed workflows, it improves as the underlying ecosystem improves: as new frontier models arrive, Sakana can fold them into Fugu's agent pool and pass the gains on to users. When a new frontier model is released publicly, they expect to spend roughly two weeks training and evaluating updated Fugu models before rolling them out.
Real-world performance: a game-building test
One early user ran a head-to-head comparison by asking both Fugu Ultra and Claude Opus 4.8 to build a "Crossy Road" clone using Three.js:
- Fugu Ultra: Completed the task in 22 minutes using ~89,000 tokens for roughly $7.32. The final game had minor logic errors, such as inverted directional turns and wonky camera angles.
- Claude Opus 4.8: Took 79 minutes, burned ~940,000 tokens for nearly $37.85, and got stuck in a retry loop requiring human intervention.
Fugu Ultra was faster and over 5x cheaper , though it shipped with bugs. That tradeoff is worth understanding before you commit to it for production tasks.
The geopolitical angle is real
Sakana is a Tokyo-based company, and the timing of this launch is not coincidental. Sakana AI targets the geopolitical and regulatory risks associated with AI sourcing. Recent export controls affecting Anthropic models like Fable and Mythos demonstrated that access to specific foundational architectures can vanish based on foreign policy decisions.
By orchestrating multiple models from different providers, Fugu essentially builds native redundancy into the AI stack. If one provider suffers an outage or faces sudden regulatory restrictions, Fugu routes around the disruption to maintain uptime. For Japanese enterprises in particular, this is a compelling pitch: frontier AI performance without betting the stack on a single foreign vendor.
Pricing and access
The flagship Fugu Ultra tier is priced at $5 per million input tokens, $30 per million output tokens, and $0.50 for cached input. These rates increase to $10, $45, and $1.00 respectively for extreme workloads utilizing context windows above 272K tokens.
Subscription plans run from $20 to $200 a month, and Sakana says subscribers before the end of July 2026 get a free second month at their tier. One important caveat: the API is not available in the EU or EEA while the company works on GDPR compliance, which immediately limits the rollout.
You can try Fugu Ultra right now on OpenRouter or directly through Sakana's console. The API is OpenAI-compatible, so dropping it into an existing integration is straightforward. For everyday tasks , short summaries, quick Q&A, simple code snippets , the standard Fugu tier is the better starting point. Save Fugu Ultra for the long, messy, multi-step work where the extra depth actually pays off.
The old assumption was that the next leap comes from the next bigger training run. Fugu Ultra points to another path: small learned systems coordinating larger specialists. Whether that path scales as reliably as raw compute remains the open question , but the benchmark numbers suggest it's worth taking seriously.