Security has two kinds of prompts, and one model rarely serves both
Ask a frontier model to audit a pull request for SQL injection, write a threat model for a new service, or turn a pentest's raw findings into a report, and it does excellent work. Ask it to explain what a malware sample in your sandbox is doing, solve a CTF binary, or decide whether a crash is exploitable, and since the Claude Opus 5.5 release on September 22, 2026 there is a growing chance it stops partway through. Anthropic's launch notes say the model ships with cybersecurity safeguards, and that when those safeguards intervened in benchmark testing the tasks were completed by Claude Opus 4.8. Since October 21, 2026, users have reported Opus 5.5 halting reverse-engineering answers mid-sentence. OpenAI and Google apply similar classifiers.
LLM Gateway puts both kinds of model behind one endpoint, so a security team can use frontier reasoning where it is allowed and a Chinese-lab model where it is needed, without a second vendor account for either.
Use the model that will finish the task
The cyber security models page lists both groups with live prices. A typical split:
| Task | Model | Why |
|---|---|---|
| Pull-request audit, threat model, report | Claude Opus 4.8, GPT-5.5 | Strongest reasoning; no classifier on defensive work |
| Daily log triage, alert summarisation | DeepSeek V4.1 Flash, Qwen3.7 Plus | $0.15–$0.40 per million input tokens at bulk volume |
| Sandboxed malware analysis, CTF, RE | DeepSeek V4 Pro, Kimi K3.1 Pro | No cyber refusal layer; 1M context; tool calling |
| Exploit triage, crash classification | GLM-5.2, Qwen3.7 Max, MiniMax M3 | Complete the analysis; strong code reasoning |
Switching is a model string. The request shape does not change.
1curl https://api.llmgateway.io/v1/chat/completions \2 -H "Authorization: Bearer $LLM_GATEWAY_API_KEY" \3 -H "Content-Type: application/json" \4 -d '{5 "model": "deepseek-v4-pro",6 "messages": [7 {"role": "system", "content": "You are a malware analyst. The sample runs in an isolated sandbox we own."},8 {"role": "user", "content": "Here is the Ghidra pseudocode for the sample'\''s entry point and its decoded strings. What does it persist, what does it exfiltrate, and which IOCs should we block?\n\n..."}9 ]10 }'1curl https://api.llmgateway.io/v1/chat/completions \2 -H "Authorization: Bearer $LLM_GATEWAY_API_KEY" \3 -H "Content-Type: application/json" \4 -d '{5 "model": "deepseek-v4-pro",6 "messages": [7 {"role": "system", "content": "You are a malware analyst. The sample runs in an isolated sandbox we own."},8 {"role": "user", "content": "Here is the Ghidra pseudocode for the sample'\''s entry point and its decoded strings. What does it persist, what does it exfiltrate, and which IOCs should we block?\n\n..."}9 ]10 }'Run the agents you already have
Security agents are MCP-driven: REA for binaries and apps, a Ghidra or Hopper bridge for native code, a sandbox wrapper for dynamic analysis, Semgrep or CodeQL for static scans. None of them care which model the agent talks to. Claude Code reads ANTHROPIC_BASE_URL=https://api.llmgateway.io; Codex, OpenCode, Cursor and Qwen Code read an OpenAI-compatible base URL. Every model on the cyber security list supports tool calling, so the agent's tools work unchanged. See the reverse engineering use case for the exact Claude Code and Codex config.
Keep long scans alive and know what they cost
A full-repository audit or an overnight log sweep is hundreds of requests. Use a model ID without a provider prefix and the gateway retries on another eligible provider when one rate-limits or errors; a model served by a single provider still fails with that provider, so prefer multi-provider models for overnight jobs. Give each pipeline its own API key — the SAST bot, the triage job, the red-team agent — and the ledger shows model, tokens, latency and dollar cost per request, grouped by key. Prompt caching stops you paying full price for the same system prompt and rule set on every call.
Stay inside the rules
Analyse systems you own or are authorised to test. Malware belongs in a sandbox, CTFs are built to be broken, and interoperability and security research carry explicit exemptions in many jurisdictions. By default the gateway stores request metadata (model, tokens, cost, latency), not prompt contents; full payload retention is a separate Enterprise setting. The upstream provider still receives every prompt, so keep client data out of prompts unless your engagement allows it, and use BYOK with your own provider keys when a contract requires it.
Get started
Create a key, point your scanner or agent at https://api.llmgateway.io/v1, and pick a model from the cyber security list for each job.