Sakana AI has introduced Fugu-Cyber (model ID: fugu-cyber-v1.0), a cybersecurity-specialized endpoint within its Fugu orchestration family. Rather than debuting an entirely new frontier model, Sakana has positioned Fugu-Cyber as a third endpoint on the Fugu orchestrator, specifically tuned for security reasoning tasks. The Fugu orchestrator was launched approximately one month earlier. The release arrives as several frontier model providers have introduced cyber-focused variants, reflecting enterprise demand for automated vulnerability analysis and detection engineering.

Sakana reports that Fugu-Cyber achieves a success rate of 86.9% on CyberGym and 72.1% on CTI-REALM, characterizing the results as comparable to cyber-focused frontier models such as GPT-5.5-Cyber and Claude Mythos Preview.

What the Benchmarks Measure

The two evaluations target opposite ends of a security workflow.

CyberGym is a UC Berkeley benchmark comprising 1,507 real-world vulnerabilities across 188 OSS-Fuzz projects. In its primary task, an agent receives a vulnerability description alongside an unpatched codebase and must write a proof-of-concept that crashes the pre-patch build but not the post-patch build. That verification step is what makes the benchmark resistant to gaming.

CTI-REALM is Microsoft's open-source detection-engineering benchmark. Microsoft curated 37 public threat reports from sources including Datadog Security Labs, Palo Alto Networks, and Splunk. An agent must map MITRE ATT&CK techniques, explore telemetry, iterate on KQL queries, and produce validated Sigma rules. Scoring spans Linux endpoints, Azure Kubernetes Service, and Azure cloud.

Together, the pair covers the spectrum from "find and prove the bug" to "turn intelligence into a detection," a framing that represents the most defensible part of Sakana's announcement.

Competitive Context on CyberGym

When CyberGym researchers published their initial results, the best agent-model pairing reached approximately 20%. Anthropic subsequently reported 83.1% for Claude Mythos Preview under Project Glasswing in April 2026. OpenAI reported 85.6% for its updated GPT-5.5-Cyber, compared with 81.8% for the standard GPT-5.5. Sakana's 86.9% therefore represents a marginal step beyond the reported frontier rather than a significant leap.

CTI-REALM presents a different picture. Microsoft's own evaluation placed the top three configurations—all Claude-based—in a range from 0.624 to 0.685. Fugu-Cyber's 72.1% would place it above that band. However, one caveat applies: CTI-REALM is scored as a trajectory reward between 0 and 1, not as a pass/fail rate. Sakana nonetheless refers to the figure as a "success rate."

Orchestration Architecture

Fugu is itself a language model, trained to read a query and dynamically construct an agentic scaffold before delegating sub-tasks to specialist models within a pool. The approach is documented in the Fugu technical report and two ICLR 2026 papers: TRINITY and the Conductor. TRINITY assigns Thinker, Worker, and Verifier roles across multiple LLMs, while the Conductor learns natural-language coordination strategies through reinforcement learning. This orchestration model lets Sakana swap underlying frontier models without retraining the endpoint itself, which matters in a landscape where base models are updated every few months.

For security work specifically, Sakana's research team emphasizes the verifier role as central. A candidate vulnerability surfaced by one agent undergoes validation by security-specialized sub-agents before any patch is proposed. Model routing remains proprietary, meaning the specific model handling each step is not visible to the user.

Access, Policy, and Pricing

Fugu-Cyber access is restricted across four dimensions.

Application required: Users must submit a form stating their intended use case and providing verified contact details. The Sakana team reviews each application manually.

Acceptable Usage Policy: The model ships under an updated AUP that prohibits offensive misuse.

Billing restriction: Usage is limited to the Token Plan. The $20, $100, and $200 subscription tiers cover Fugu and Fugu-Ultra only.

Geographic limitation: The Fugu API is not available in the EU or EEA while Sakana works toward GDPR compliance.

Pricing is fixed at $6 per million input tokens, $36 per million output tokens, and $0.60 per million cached input tokens. All three rates double above a 272K-token context window. Each rate is exactly 1.2 times the Fugu-Ultra rate, representing a flat 20% premium for the cybersecurity endpoint. Since lengthy codebase analyses frequently exceed 272K tokens, the doubled tier is not an edge case.

Fugu-Cyber launched on July 21, 2026. Both reported scores are self-reported and have not been independently replicated. Self-reported benchmark results are common practice across the AI industry, making external verification through community replication or third-party audits the key milestone to watch for. Model weights are not released. Sakana's stated position is that a capable API combined with human security expertise outperforms the API alone.