Fugu Cyber: Sakana’s Access-Gated Security Model, Explained

Fugu Cyber is the member of the Fugu family you can’t just switch on. It’s Sakana’s security-specialized model — built for security analysis, vulnerability research, and threat investigation — and it sits behind an access-request form, a manual review, and a mandatory billing mode. This guide covers what it is, how the access process works, how it’s billed, and where it fits in a defensive security practice. Verified against Sakana’s official pages on August 19, 2026.

What Fugu Cyber is

Architecturally, Fugu Cyber is Fugu: a trained orchestrator that assembles agents from a pool of frontier models behind one API, with the family’s 1,000,000-token context window, text and image input, and high/xhigh reasoning-effort levels. What changes is the specialization — Sakana describes it as “specialized for cybersecurity reasoning tasks,” aimed at security analysis, vulnerability research, and threat investigation.

That’s a meaningful niche. Security work is long-context by nature (logs, packet captures, codebases, threat reports), verification-heavy (a missed detail is the whole failure), and multi-step — exactly the shape of problem the family’s Thinker/Worker/Verifier architecture was built around. A security-tuned orchestrator that can read a large evidence corpus in one request is a genuinely different tool from a general chat model with a security prompt.

The access process: gated on purpose

Security-capable AI is dual-use by definition, and Sakana treats it that way. Getting Fugu Cyber involves three documented requirements:

1. An access-request form detailing your intended use case, with verified contact information.
2. Manual review by Sakana before your endpoint access is enabled — this is not an instant toggle.
3. Pay-as-you-go billing mode on your API key. Fugu Cyber is available only through pay-as-you-go billing — no subscription plan includes it — and the model won’t even appear in your key’s model list unless the billing mode is set accordingly.

Plan for the review in your timeline: this is a procurement, not a signup. And treat the use-case statement seriously — Sakana’s terms attribute all usage on a key to its owner, and access granted for one purpose is not a blank check.

Billing: PAYG only, pricing via sales

Fugu Cyber publishes no rate table. Sakana’s pricing page says to contact the sales team for usage and pricing details — so budget conversations happen before access, not after. Two knowables regardless (as of August 2026): billing is pay-as-you-go only, funded by the same prepaid credits as the rest of the platform (six-month expiry — see the pricing guide), and like Fugu Ultra, Cyber reports orchestration usage that is billed at standard rates on top of visible tokens. Whatever your negotiated rates, budget from total billed tokens.

Reported benchmark positioning

Sakana’s June 2026 evaluation reports Fugu Cyber at 86.9% on CyberGym and 72.1% on CTI-REALM — benchmarks covering security task execution and cyber threat intelligence reasoning respectively — against baselines drawn from leading frontier models. The usual caveats apply doubly in security: these are vendor-reported numbers from a point-in-time evaluation, and benchmark performance is not an assessment of fitness for your threat model. Verify against your own workload before relying on it.

Where it fits: defensive use cases

The legitimate home for a tool like this is the defensive side of a security practice, and that’s the only side we cover:

• Security analysis and code review — auditing your own codebases and configurations for vulnerable patterns before an attacker finds them, with the long context window fitting entire services into one review.
• Vulnerability research on assets you’re authorized to test — triaging findings, assessing exploitability and impact, prioritizing remediation.
• Threat investigation — working through logs, alerts, and incident evidence; correlating indicators; drafting the incident timeline a human analyst then verifies.
• Threat-intelligence work — the CTI-REALM benchmark’s territory: digesting and reasoning over threat reports at scale.

What we won’t cover, here or anywhere on this site, is offensive tooling or techniques. Beyond our own policy, the practical point stands on its own: Sakana reviews your stated use case, attributes all usage to your key, and can terminate access — a defensive practice with authorization paperwork in order is what this product is for.

Should your team pursue access?

A quick filter. Fugu Cyber is worth the process if your security team already has AI-assisted workflows and hits the ceilings of general models on security-specific reasoning; if long-context evidence analysis (whole codebases, full incident corpora) is your bottleneck; or if you’re evaluating specialized models against your SOC’s real tickets. It’s premature if you haven’t yet tried the general models on the same tasks — Fugu Ultra needs no access request, handles plenty of security-adjacent reasoning, and can be tried in minutes in our playground — or if your budget needs published rates to plan against.

Context for the rest of the family: What is Sakana Fugu? and Fugu Ultra explained. Costs across the platform: the pricing guide.