OrcaCyber Zero 1.5 Claims 100% Cybench, 1M Context

OrcaRouter has released OrcaCyber Zero 1.5, a post-trained cybersecurity model built for vulnerability reproduction, exploit development, penetration testing, and authorized security engineering. The successor to OrcaCyber Zero 1.0, which shipped on September 17, 2026, pairs a 1,000,000-token context window with a 128,000-token maximum output. Orca lists the model through an OpenAI-compatible hosted endpoint at $3.00 per million input tokens and $7.50 per million output tokens; cache reads cost $0.75 per million tokens.

Access is gated to Orca's Security Research tier. Orca requires an active engagement, a passkey, and accepted terms, targeting trusted security researchers, red teams, and authorized tests. The model accepts text input and text output, with native function calling and structured outputs.

Orca's benchmark results, last evaluated on October 10, 2026, are self-reported. The vendor reports 100% on Cybench under unrestricted agent execution, covering 39 of the suite's 40 professional CTF tasks. For CVE-Bench, Orca reports 95.8% on a 24-task evaluable subset — not the full 40 critical web CVEs — which works out to 23 solved tasks. The model also reports 93.9% on HumanEval+ and 76.5% on SWE-bench Pro V2, its lowest published score. Orca has not published a CyberGym score for Zero 1.5. The predecessor, OrcaCyber Zero 1.0, reported 98.07% pass@1 on CyberGym (1,478 of 1,507 tasks), but that figure came through the OrcaCyber Harness rather than a single-model call.

OrcaRouter's comparison table, with competitor figures supplied by each vendor, breaks down as follows:

OrcaRouter's live production metrics list a p50 time-to-first-token of 3.22 seconds, a p95 TTFT of 9.46 seconds, output speed of 73.9 tokens per second, and an error rate of 0.14% across 3,302.5 million tokens processed over seven days. Model weights, quantization schemes, and the underlying compute hardware remain undisclosed.

At $3.00 input and $7.50 output, OrcaCyber Zero 1.5 undercuts frontier offensive alternatives. MarkTechPost reports that Anthropic lists Claude Mythos Preview at $25.00 per million input and $125.00 per million output after platform credits. OrcaRouter's cost calculator estimates a 10 million token monthly workload with 70% input and 30% output at $43.50 at list rates, or $35.63 with prompt caching; the same workload on Mythos Preview is $550.00.

According to AI Mastery's pricing index as of September 14, 2026, Orca's $3.00 input price matches Claude 4.6 Sonnet ($3.00 input / $15.00 output) and sits below GPT-5.6 Sol ($4.00 / $20.00) and Claude Opus 5 ($5.00 / $25.00). The result is security-specialized 1M-context pricing at the level of mid-tier general models.

AI Mastery analysis

The benchmark scope is the main due-diligence risk. Orca's 100% Cybench and 95.8% CVE-Bench results look strong, but both scores leave out hard parts: one of 40 Cybench tasks and 16 of 40 CVE-Bench challenges fall outside the reported subsets. Without a reproduction harness — Orca has not published one — engineering teams cannot assume the accuracy transfers to messier legacy codebases.

A 1M-token context also does not guarantee structured agent execution. Security agents routinely misattribute logic when cross-referencing authentication boundaries with third-party service endpoints. As we examined in AI agent security fails at shared surfaces, not model logic, monolithic context often shifts the failure point from detection to tool execution and permission handling.

Deploying OrcaCyber Zero 1.5 as an automated exploitation engine requires strict output isolation. Because the model is tuned for exploit development and attack paths, CI/CD testbeds should isolate runner environments so speculative exploit traffic cannot reach staging assets.

The release is another case of post-training specialization undercutting general frontier prices in security verticals. If long-context execution holds up on uncurated vulnerability testbeds, specialized offensive models will take workloads that general labs have priced out of autonomous use.

Sources

• Primary: OrcaCyber Zero 1.5 API — Pricing & Benchmarks · OrcaRouter
• Reported by: MarkTechPost: OrcaRouter Releases OrcaCyber Zero 1.5 Cybersecurity Model With 1M Context
• Competitor prices: AI Mastery token calculator, figures as of 2026-09-14