OrcaRouter has released OrcaCyber Zero 1.5, a reasoning model built specifically for cybersecurity work: finding vulnerabilities, developing exploits, running penetration tests and auditing systems. It succeeds OrcaCyber Zero 1.0 and was developed by Continuum AI.
The model's headline feature is context length: a one-million-token window, with output capped at 128,000 tokens. It supports native function calling and structured outputs, which matters for a model meant to feed results into other security tooling rather than just produce prose.
Access is restricted. OrcaCyber Zero 1.5 sits behind a gated "Security Research" tier that requires an engagement, a passkey and accepted terms before anyone can query it through its OpenAI-compatible API. Pricing is $3 per million input tokens and $7.50 per million output tokens.
OrcaRouter's own benchmark figures are striking: 100% on Cybench under unrestricted agent-execution conditions, 95.8% on an evaluable subset of CVE-Bench, 76.5% on SWE-bench Pro V2 and 93.9% on HumanEval+. These numbers come entirely from the vendor; CVE-Bench, in particular, was run against only 24 evaluable tasks, a small sample for a 95.8% claim. No independent lab has yet reproduced these results.
The gating makes sense given the model's purpose. A tool this capable at finding and reproducing exploits is also a tool that could cause real damage if it reached the wrong hands, so OrcaRouter is trading open access for accountability. Whether the benchmark claims hold up under outside scrutiny is a separate question, and one the security research community will likely test quickly.
Source: MarkTechPost
