Browser
AI models for cybersecurity
— what they actually do well.
A vendor-neutral comparison of frontier models for security work.
What is Claude Mythos?
Claude Mythos is a generative AI model from Anthropic, released as a gated research preview.
In Anthropic's testing it discovered a 27-year-old bug in OpenBSD and achieved full control-flow hijack on 10 fully-patched targets.
What is Project Glasswing?
Glasswing is Anthropic's restricted-access program for Mythos, gated to ~12 launch partners plus ~40 additional organisations.
Comparison table — frontier models for security work
Numbers below are vendor-stated specs as of May 2026.
| Model | Context | Pricing (in/out per 1M) | Access | Best at (security) |
|---|---|---|---|---|
| Claude Mythos Preview gated |
1M / 128K out | $25 / $125 | Glasswing only | Multi-step vuln research · exploit chaining · PoC generation · ~73% expert CTF |
| Claude Opus 4.7 closed |
200K / 64K out | $15 / $75 | Public API · Claude Code | Code review · long agentic flows · structured findings · architectural reasoning |
| Claude Sonnet 4.6 closed |
200K / 64K out | $3 / $15 | Public API · Claude Code | SAST-style code review · CWE classification · cost-sensitive bulk analysis |
| GPT-5.5 closed |
400K / 128K out | $10 / $40 | Public API · ChatGPT | General reasoning · tool-use · structured JSON output · CTF intermediate level |
| Gemini 2.5 Pro closed |
2M / 64K out | $2.5 / $10 | Public API · Google AI Studio | Very long context · monorepo whole-codebase grep · cheap bulk analysis |
| Llama 3.3 70B open weights |
128K / native | self-hosted | HuggingFace · Together · Groq | Self-hosted air-gapped audits · no data leaves perimeter · BYO infrastructure |
| Mistral Large 2 open weights |
128K / native | self-hosted | HuggingFace · Mistral La Plateforme | European-data-residency audits · self-host EU deployments |
CTF benchmarks don't translate 1:1 to real-world exploitation — production systems have WAFs, EDR, defenders.
Which model for which security task?
Claude Sonnet or Opus
Cheap, structured CWE output, long context.
Claude Mythos (if you have access)
Out of reach for 99% of teams. Opus or GPT-5.5 is the 80%-good alternative.
Gemini 2.5 Pro
2M token context fits a whole mid-sized monorepo. Cheapest per-token.
Llama 3.3 70B or Mistral Large 2
Open weights, self-hosted, no data leaves your perimeter.
Cost & access — who can actually use what
Sonnet 4.6 at $3/$15 is the workhorse. Opus jumps to $15/$75. Mythos at $25/$125 is the most expensive frontier model on the market.
Our /mythos harness uses GPT-4o-mini and caps each run at ~$0.30-1.50.
Air-gapped? Point /mythos at any OpenAI-compatible self-hosted endpoint — OpenMythos, Llama 3.3 70B via Ollama, or Mistral La Plateforme.
FAQ
Skip the access waitlist.
Browser's /mythos harness implements the same architectural pattern Cloudflare published — parallel narrow hunters + adversarial validator + taint-trace — using GPT-4o-mini at ~$0.30-1.50 per run. No Glasswing required.
Run the Browser harness → Or browse our advisories