Sakana's Fugu Ultra Beats GPT-5.5 and Claude by Orchestrating Both

Sakana AI's Fugu lands in OpenCode, the same tool used to build it — a multi-agent orchestration system that beats frontier models by coordinating them

·
·
·
Sakana's Fugu Ultra Beats GPT-5.5 and Claude by Orchestrating Both
  • Fugu joins OpenCode: Sakana AI's multi-agent orchestration system is now available inside OpenCode, the same tool used to build it.
  • Not a new model, a conductor: Fugu is a small trained orchestrator (TRINITY ~0.6B, Conductor 7B) that dynamically routes tasks across GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro, and others.
  • Frontier-beating benchmarks: Fugu Ultra scores 73.7% on SWE-Bench Pro (vs. 69.2% for Claude Opus 4.8) and 82.1% on TerminalBench 2.1, though GPT-5.5 leads on long-context recall (MRCRv2).
  • OpenAI-compatible API, two tiers: Fugu (low-latency, everyday) and Fugu Ultra (max quality, 1M context) — swap with one line of code, no SDK migration needed.
  • Pricing: Subscriptions from $20–$200/month; pay-as-you-go at $5 input / $30 output per 1M tokens for Fugu Ultra. Not available in EU/EEA.
  • Geopolitical pitch: Launched days after Anthropic's export-control-driven access revocation, Fugu positions itself as a hedge against single-vendor dependency.

Sakana Fugu is now available inside OpenCode, the open-source coding agent that Sakana's own team used while developing Fugu's orchestration system. The integration is more than a distribution milestone: it closes a loop where the tool that helped verify Fugu's multi-agent behavior is now powered by Fugu itself. But to understand why that matters, you need to understand what Fugu actually is.

Not a model. A conductor.

Sakana Fugu is a multi-agent system that behaves like a single model. You send a request to one endpoint, and Fugu decides how to handle it: solving it directly when that is enough, or assembling and coordinating a team of expert models when a task calls for more. It manages model selection, delegation, verification, and synthesis internally, so the complexity of a multi-agent system never reaches your code.

Fugu owns no frontier weights of its own. It is a small orchestrator that learns to combine existing models like GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro, and a bench of open-source LLMs into one collective. The routing is proprietary, and it never even tells you which underlying model answered.

The research behind the curtain

The system rests on two ICLR 2026 papers. TRINITY is a roughly 0.6B-parameter coordinator, evolved with CMA-ES (an evolutionary optimization algorithm), that assigns Thinker, Worker, and Verifier roles across a pool of much larger worker models. Conductor is a 7B model trained with reinforcement learning to discover natural-language coordination strategies, and it can call itself recursively to scale compute at test time. The orchestrator is tiny; the heavy lifting is delegated to a swappable pool of frontier models behind the scenes.

The key insight from both papers is that coordination should be learned, not hand-wired. Most multi-agent systems require developers to define which model handles which task. Fugu learns this. The orchestrator model is trained to decide when to delegate, how agents should communicate, and how to combine their outputs into a single answer. In practice, this means you never write a LangChain graph or a CrewAI workflow , Fugu figures out the topology on its own.

Two tiers, one API

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. Fugu picks a single best worker per query, so its latency is comparable to a direct call to a frontier model.
  • 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. Fugu Ultra supports a maximum context window of 1,000,000 tokens.

How it actually performs

The headline claim is that Sakana Fugu Ultra matches Anthropic's Fable 5 and Mythos Preview on engineering, scientific, and reasoning benchmarks, without requiring access to either of those export-controlled models. The numbers back it up across most tasks:

BenchmarkFuguFugu UltraClaude Opus 4.8Gemini 3.1 ProGPT-5.5
SWE-Bench Pro59.073.769.254.258.6
TerminalBench 2.180.282.174.670.378.2
LiveCodeBench92.993.287.888.585.3
GPQA-D95.595.592.094.393.6
Humanity's Last Exam47.250.049.844.441.4
MRCRv2 (long-context recall)86.693.687.984.994.8

GPT-5.5 wins MRCRv2 with 94.8%, compared to Fugu Ultra at 93.6%. MRCRv2 tests long-context recall , specifically whether a model can retrieve specific information from very long documents. This is the one benchmark where Fugu Ultra does not lead, which is worth noting for teams whose primary use case involves retrieving specific facts from large documents rather than multi-step reasoning or code generation.

What users are actually doing with it

The beta ran with close to 500 early users. The patterns that emerged are telling:

  • Code review: Users report Fugu Ultra surfacing 20+ issues where other tools flagged only 3.
  • Autonomous ML research: An AI agent autonomously improved a small GPT's training recipe using AutoResearch (Karpathy et al.), iteratively editing training code and running experiments over ~14 hours on a single H100 GPU , 123 experiments total. Fugu Ultra finished with the best mean BPB across all runs.
  • Security assessment: Given one scoped instruction, Fugu drove a full end-to-end security assessment covering recon, XSS/SQLi checks, auth review, and a clean report with evidence , staying inside scope throughout.
  • Patent and literature research: A researcher mapped a patent landscape across ~20 papers and several patents in hours instead of 3-4 days, including connections they would not have spotted manually.
  • Paper reproduction: From a single prompt, Fugu worked autonomously for nearly four hours , reading, implementing, training, evaluating, and analyzing gaps.

The geopolitical angle Sakana is leaning into

Sakana CEO David Ha, formerly of Google Brain, positioned Fugu as a more reliable option for enterprise workflows than any single AI model provider, in the wake of Anthropic's move to revoke public access to its most powerful models, Claude Mythos 5 and Claude Fable 5, following a U.S. government export control order.

If a single provider restricts access, Fugu dynamically routes around the disruption. Over time, Sakana Fugu will naturally grow by incorporating newer, more efficient models, including their own. By orchestrating the world's models, they are delivering what they call the realistic, resilient blueprint required for AI sovereignty.

Adopting Fugu to reduce vendor dependency does add a new dependency , on Sakana's orchestrator. That can still be net-positive, but it is a trade, not an escape. The more honest framing: Fugu is a real hedge against provider-level disruption, but it introduces its own concentration risk at the orchestration layer.

Getting started

Although Sakana Fugu is a multi-agent system, it can be used like a standard LLM through the Sakana API, similar to GPT, Gemini, or Claude. The API supports both Chat Completions and Responses endpoints, with the Responses API providing broader compatibility with modern tooling.

Calling Fugu Ultra from Python is a one-line swap from any existing OpenAI client:

python
from openai import OpenAI
client = OpenAI(
    base_url="https://api.sakana.ai/v1",
    api_key="YOUR_SAKANA_API_KEY"
)
response = client.responses.create(
    model="fugu-ultra",
    input="Reproduce the key experiment from this paper: ...",
    timeout=120.0,
)
print(response.output_text)

For OpenCode specifically, the integration is a config drop-in. Sakana provides a direct, one-line install command to integrate the model natively into Codex, allowing developers to leverage multi-agent orchestration right inside their code editor. Both fugu and fugu-ultra accept two reasoning effort levels: high and xhigh.

Pricing

There are two billing modes. Subscriptions cover both Fugu and Fugu Ultra:

  • Standard: $20/month , lightweight daily use
  • Pro: $100/month , 10x Standard allowance, for regular working sessions
  • Max: $200/month , 30x Standard, for heavy long-running workloads

On the pay-as-you-go plan, standard Fugu charges a dynamic rate based on the specific underlying models activated, whereas Fugu Ultra uses a fixed pricing structure starting at $5 per million input tokens and $30 per million output tokens. Fugu Ultra supports a maximum context window of 1,000,000 tokens, though standard API pricing applies to the first 272,000 tokens, with rates increasing for contexts that exceed that threshold. The service is not yet available in the EU/EEA.

What this changes

The bigger signal is not this one product. It is that "model orchestration as a product" is now a real category, sitting alongside aggregator-style routing, DIY frameworks like LangGraph and CrewAI, and in-harness dynamic workflows from the model vendors themselves.

Sakana Fugu is the most concrete bet yet that the next gains come from coordinating models rather than scaling one , and the research behind it (TRINITY, the Conductor) is real, peer-reviewed, and on-brand for a lab that has spent years on collective-intelligence methods. Delivering all of that behind a single OpenAI-compatible API is a genuinely useful piece of productization. The OpenCode integration makes that bet accessible to anyone already using a terminal-based coding agent , no new SDK, no new mental model, just a different endpoint.

Trending
  • No trending articles

Comments

avatar

Next Reads