3. OrcaRouter — Native Cyber Models

3.1 Identity

• Name: OrcaRouter. Site https://www.orcarouter.ai/ ↗, docs https://docs.orcarouter.ai ↗, models https://www.orcarouter.ai/models ↗, pricing https://www.orcarouter.ai/pricing ↗, offers https://www.orcarouter.ai/offers ↗.
• What: “One OpenAI-compatible AI gateway — adaptive routing, load balancing, guardrails, agent firewall, observability across 200+ models.” Grades prompts in <1ms, $0 markup.
• Operator: CONTINUUM AI PTE. LTD., Singapore, 21 Merchant Road, legal@orcarouter.ai. Trade identity Continuum AI Corp, San Francisco, founded 2026, 5-6 people. Launch PR May 8 2026 (100+ models → 200+ now).
• Footprint: GitHub Continuum-AI-Corp/OrcaRouter-Lite (1.7k stars, 270 forks, MIT, 403 tests), HuggingFace orcarouter (#10 across 3M+ models claim, 1.5M downloads/30d), Ollama orcarouter, X @OrcaRouter, Discord, LinkedIn 1.6k followers, arXiv RouterArena 75.54% #2.
• Scale (vendor-reported): 203 models per LLMScape Sep 28 2026, max 1.1M ctx, 162 text/multimodal, 179 with pricing.

3.2 Cyber Models (Confirms >1)

3.2.1 OrcaCyber-Zero-1.0 — native offensive, gated closed beta

• Page: https://www.orcarouter.ai/models/orca/orcacyber-zero-1.0 ↗
• Function: coding model post-trained for vuln analysis, vuln reproduction, red-teaming, offensive research. Frontier-tier in OrcaCyber Harness. Best for reproducing/triaging real vulns across large OSS codebases, root-cause, red-team tooling, long-context repo understanding with agentic function calling.
• Specs: 1M ctx, 128K max output, Tools+JSON+Reasoning, function calling, structured outputs, configurable reasoning. p50 TTFT 972ms, p95 5.31s, 196 tok/s, 0.59% error, 1895M tokens/7d traffic.
• Benchmark (vendor): CyberGym Level 1 pass@1 98.07% (1,478/1,507) from 188 OSS projects, evaluated 2026-09-17.
• Pricing: $3.00/M input, $5.00/M output, $0.30/M cache read. Calculator: 10M/mo 70% in = ~$36/mo ($26.55 cached).
• Access: “Offensive-capability model · engagement + passkey + terms” + Enable access button. Trusted researchers/red-teams/authorized testing only. Per-use-case approval: “arrives with a target attached: component you maintain, vuln class, engagement this quarter. General chat does not queue ahead.”

3.2.2 Uncensored / abliterated open-weight family — 2nd cyber-capable option

• HF org: “research-oriented and uncensored variants intended for authorized security research, red-team testing…” hosted at https://www.orcarouter.ai/models?q=uncensored ↗.
• Ollama: Qwen3.8-27B-Uncensored (tensor-level abliterated, 0% over-refusal XSTest, 0-6% refusal A/B, no measurable loss, 262K ctx, 199.6K pulls); Qwen3.8-Flash-Next-Uncensored (MoE ~177B total / ~6B active, 262K, MLX 4/6/8-bit, Research use only, 12.2K pulls).
• Hosted router IDs: obsidian/qwen3.8-27b ($0.40 in / $4.21 out snapshot, 262K), qwen/qwen3.8-27b ($0.33/$2.40), qwen/qwen3.8-27b-free ($0, 66K). HF quant orcarouter/OrcaSAQ-2-27B (55GB→12GB, 93.2% Top-1, 70.0 SWE-bench Verified, 58.4 Terminal-Bench 2.1).
• Note high churn — verify model ID before engagement.

3.2.3 Security-adjacent — OrcaVerify text classifier

orca/orcaverify-text1.0 + free variant — AI-vs-human detector (AI_GENERATED / AI_ASSISTED / HUMAN / ABSTAIN + calibrated prob + paragraph localization). $2.00/M in, p50 177ms. For UGC moderation, academic integrity, not offensive.

3.2.4 Routable dual-use frontier models

Any of 200+ for pentest assistance (code review, exploit analysis, patch validation) subject to upstream AUPs. Examples: openai/gpt-5.6-sol ~$2-4/M in, 1.1M ctx; openai/gpt-5.5 $5/$30; anthropic/claude-opus-4.8 $5/$25; google/gemini-3.8-flash $0.75/$3.75 promo thru end-2026 → doubles Jan 1 2027; deepseek/deepseek-v4-flash $0.15/$0.29; qwen/qwen3.7-flash $0.03/$0.13; specials orcarouter/auto, orcarouter/free ($0).

3.2.5 NOT callable via OrcaRouter (do not misclaim)

Gemini 3.8 Flash Cyber (Google Fairwind-gated, CyberGym 86.2%), GPT-5.6-Cyber (OpenAI Daybreak Red-gated, 95.0% Advanced Cyber Completion), MAI-Cyber-1-Flash, Gemini 3.5 Flash Cyber, Claude Mythos 5 — discussed on Orca blog for comparison only.

3.3 Pricing & Plans

• Model: zero markup — pay provider published rate. Live prices refresh every 60s.
• Plans: Hacker Free forever (200+ models, auto-failover, basic dashboard, 10 keys, 0% markup). Team Custom on live page (10 seats, compliance enforcement/reports, unlimited keys, priority support) — meta description still says $49/mo (flag as stale). Enterprise Custom (unlimited seats, beta, dedicated infra, 99.99% SLA). Prepaid top-ups + Stripe subs, per-token deduction, non-refundable, cancel end-of-period.
• Free tier rotating $0: deepseek-v4-flash-free, glm-5.3-flash-free, hy3-free, orcaverify-free, orcarouter/free. Voucher/student/hackathon credits via /offers (2-click claim).

3.4 API

• Endpoint https://api.orcarouter.ai/v1 ↗, keys sk-orca-*. Fully OpenAI-compatible: Chat, Responses, Embeddings, Images, Audio, SSE + [DONE]. Swap base_url only.
• Fallback: extra_body={"models":["openai/gpt-5.5","anthropic/claude-opus-4.8"],"route":"fallback"}; auto: model="orcarouter/auto". Resolved model in x-orca-resolved-model, cache x-orca-cache: HIT/MISS.
• Native Anthropic + Gemini ingress, same routing/cache/analytics. Official MCP server, Dify plugin, Vercel AI SDK, GET /v1/models discovery, budgets/roles, prompt versioning/caching, guardrails+agent firewall. Self-host OrcaRouter-Lite (model="auto", BYOK, SQLite, localhost:8000/v1).

from openai import OpenAI
client = OpenAI(base_url="https://api.orcarouter.ai/v1", api_key="sk-orca-...")
r = client.chat.completions.create(model="orca/orcacyber-zero-1.0",
  messages=[{"role":"user","content":"Analyze this diff for RCE, authorized test ID ..."}])

3.5 Legal / Guardrails

• ToS §5 bans violating law, CSAM, IP theft, “create malware, stalkerware, or other software designed to harm”, harassment, impersonation, limit circumvention, scraping. Upstream AUPs also bind you (OpenAI, Anthropic, Google); violation upstream = violation OrcaRouter; abuse metadata may be shared upstream. Unauthorized offense prohibited; authorized testing via gated cyber is allowed path.
• Cyber gating = guardrail: closed beta, per-use-case approval, engagement+passkey+terms.
• Controls: guardrails + agent firewall all plans; Team compliance reports; Enterprise data residency, audit, SLA. Trust Center maps HIPAA/PCI/GDPR/SOC2/AI Act/NIST/OWASP/ISO 27001/42001 (capability mapping, not audit; reports under NDA). Regions SG/US/JP (workspace region ≠ inference pinning). You retain rights, no training on content. Outputs may be inaccurate — evaluate before reliance. Jurisdiction Singapore (SIAC).