Sakana Fugu Cyber: 86.9% on CyberGym, How to Apply

Sakana Fugu Cyber Puts a Pool of Agents Behind One Endpoint

Sakana AI's Fugu Cyber, announced on July 21, 2026, is an update to the company's Fugu orchestration model, released as a new API endpoint for cyber defense. Orchestration here means a coordinating layer that decides which agent handles which part of a task, so the caller only ever sees one interface.

Fugu Cyber is an update to Sakana Fugu from Sakana AI

The announcement calls Fugu Cyber an update to the original Sakana AI Fugu model, and it keeps the same design: a multi-agent system that behaves like a single model. You send a request to one endpoint, and the system picks from a pool of specialized agents for complex, multi-step tasks.

Sakana AI adds that the pool comes without the risk of depending on a single vendor. What is new is the target: the console models page says Fugu Cyber is specialized for cybersecurity reasoning, and lists security analysis, vulnerability research and threat investigation as its work.

Fugu Cyber vs. Sakana Fugu Ultra

Fugu Cyber is the cybersecurity entry in the console's Fugu lineup, alongside Fugu, Fugu Ultra and Fugu Max. The "Fugu lineup in the console" table condenses the descriptions given there.

Fugu lineup in the console

The practical difference from Fugu Ultra is scope and access. The console describes Ultra by task difficulty and Fugu Cyber by domain, and it ties Fugu Cyber to an application and pay-as-you-go billing.

CyberGym and CTI-REALM Benchmark Scores: 86.9% and 72.1%

Sakana AI puts Fugu Cyber at 86.9% on CyberGym and 72.1% on CTI-REALM. The two numbers come from different tests, and Sakana AI is the only source for both.

What CyberGym and CTI-REALM measure

CyberGym checks whether an agent can work through a complex codebase and confirm that a real-world vulnerability is actually there. CTI-REALM checks whether an agent can take a raw threat intelligence report and produce a working detection rule. One test looks at code, the other at written reports, which is why the two percentages answer different questions.

How Sakana Fugu Cyber compares with Mythos Preview

The announcement calls the results comparable to GPT-5.5-Cyber and Mythos-Preview, and that named pair is the whole comparison. It describes them as leading cybersecurity-focused frontier models. Neither the announcement nor the console gives a score for either one or says which versions were used, so read the line as Sakana AI's positioning, not as a side-by-side table.

Where the CyberGym benchmark numbers come from

Both figures appear on two pages, and both pages belong to Sakana AI. The announcement carries the sentence quoted above, and the console models page repeats it word for word under the Fugu Cyber (v1.0) entry. So the second page is not a separate confirmation; it is the same vendor restating its own result.

Neither page says how the benchmark runs were configured, and neither cites a result from a third party. For a security team, that makes 86.9% and 72.1% a reason to run its own trial on its own code and reports, not a figure to plan around.

Why Sakana AI Says Fugu Cyber Needs More Than an API

The announcement does not stop at the scores: it says a high score is only the beginning of the story. Its section titled "The Reality Check on Frontier Cyber Capabilities" makes the case. That case comes from a launch post for Sakana AI's own product, so each point is the company's claim, not an outside finding.

Cyber defense AI and Fugu Cyber: false positives and verification

The announcement says a raw model deployed in isolation will produce false positives, so its findings need checking before anyone acts on them. Such a model needs a harness, meaning the software and process wrapped around a model to feed it the right context and check what it returns. When a system surfaces a possible vulnerability, cybersecurity sub-agents and human review should confirm it would trigger in a real environment before anyone proposes a patch.

The announcement cites a Nikkei Digital Governance report (in Japanese) and uses Anthropic's Mythos as its example: access to a frontier model, it says, does not by itself solve enterprise security. Large organizations, it adds, often struggle to put such tools to work without specialized internal talent and deep integration with their own source code.

For anyone evaluating cyber defense AI, this changes the question from "what does it score?" to "what checks the output?" By the announcement's account, the API alone does not answer that.

Sakana AI's Applied Enterprise team and Sakana Fugu Cyber harnesses

Sakana AI says its Applied Enterprise team is currently building those harnesses. The team works with major Japanese institutions on the harnesses and workflows needed to run these models in production, including automated vulnerability verification. The announcement ends by pointing readers who want enterprise solutions to its Applied team.

That marks where the API stops. Fugu Cyber is the model behind the endpoint; by Sakana AI's account, the verification workflow around it is work the company is doing with major Japanese institutions.

How to Get Sakana Fugu API Access and How Billing Works

As of October 1, 2026, Fugu Cyber is available at the sakana.ai/fugu endpoint, but only after your application is approved. Approval is a manual review, and the billing mode on the API key decides whether the model even appears.

Applying for Fugu Cyber access from Sakana AI

You apply through an access request form that asks for your intended use case and contact information that can be verified. Sakana AI's team then reviews and approves each application by hand before the endpoint opens to you.

Once approved, you can call it through familiar API formats. According to the console models page, Fugu supports the OpenAI-compatible Chat Completions, Responses and Models APIs, plus the Anthropic-compatible Messages API.

The page recommends the Responses API for generation requests. The model ID is fugu-cyber, which defaults to fugu-cyber-v1.0.

Sakana Fugu pricing: Token Plan or pay-as-you-go for Fugu Cyber

Fugu Cyber's billing is worded differently on the two Sakana AI pages, and the console's version decides how the API key must be set. The announcement names the Token Plan, and the console names pay-as-you-go billing.

The console adds one detail that affects setup: the models endpoint returns fugu-cyber only under pay-as-you-go billing. If your key is on another billing mode, the model will not show in the list.

On cost, the console says Fugu Ultra and Fugu Cyber return usage fields that separate orchestration work from the visible input and output. Orchestration tokens are billed at the same rate as standard input and output tokens and count toward the request's final price. Neither page lists a price for Fugu Cyber, so check with Sakana AI before you budget.

Usage policy: no offensive use of Sakana Fugu Cyber

Fugu Cyber is released under an updated Acceptable Usage Policy that prohibits offensive misuse. Sakana AI ties this to the sensitive security workflows Fugu Cyber handles and says it is committed to a safe and responsible release. Because the form asks for your intended use case and each application is reviewed, it makes sense to describe your defensive work specifically.

Summary: Fugu Cyber scores are Sakana AI's own; access needs approval

Fugu Cyber is Sakana AI's update to its Fugu orchestration model for cyber defense: one endpoint in front of a pool of specialized agents. It is pitched at security analysis, vulnerability research and threat investigation. The scores, 86.9% on CyberGym and 72.1% on CTI-REALM, are Sakana AI's own, and the announcement and the console models page carry the same sentence without a run configuration.

The announcement itself says a high score is only a start and that raw models deployed in isolation will produce false positives. Findings need verification by cybersecurity sub-agents and people before a patch is proposed.

To use the model, you apply with a use case and verified contact information and wait for a manual review. The announcement says Token Plan, the console says pay-as-you-go only, and neither page lists a price.