Anthropic Built Opus 4.7 for Cyber Defense. It Scored Last.
Anthropic released Claude Opus 4.7 in April with explicit cybersecurity marketing: the Cyber Verification Program, automated detection of high-risk prompts, and product copy describing “cyber-grade safeguards.” It was the first Claude built for security work. On Simbian’s Cyber Defense Benchmark — 1,206 real attack-log investigations across 13 of 14 MITRE ATT&CK tactics and 105 chained kill-chain procedures — Opus 4.7 finished last.
The Leaderboard
| Rank | Model | Coverage | Median Cost | Tactics Passing |
|---|---|---|---|---|
| 1 | Claude Opus 4.6 | 44.5% | $2.71 | 7 of 13 |
| 2 | Claude Sonnet 4.6 | 36.9% | $2.21 | 2 of 13 |
| 3 | Claude Opus 4.8 | 32.7% | $1.69 | 1 of 13 |
| 4 | Claude Opus 4.7 | 30.9% | $1.65 | 0 of 13 |
GPT-5.5 scored 26.4%, GPT-5 scored 17.4%. Every Anthropic model on the board outranked every OpenAI and Google model. The best LLM for SOC work today is the one released five months before the cyber-specific one.
Why the Cyber Model Performs Worst
Simbian’s explanation is structural. Opus 4.7 optimizes for speed and cost: its median investigation runs $1.65 at 243 seconds — the fastest and cheapest Claude tested. For cybersecurity defense work, that is a liability. SOC reasoning requires hypothesis generation under heavy noise, MITRE-coverage planning, and iterative queries across 100K+ events per environment. Investigation depth tracks investigation time. The model designed to move fast through security tasks moves too fast to do security work well.
The Cyber Verification Program safeguards block offensive prompts as designed. They do not add defensive reasoning skill. The benchmark measures defense, not safety posture, and those are not the same thing.
Opus 4.8 does not recover the gap. At 32.7% coverage, it passes 1 of 13 MITRE tactics and trails Opus 4.6 by 12 points. Newer models optimize for general benchmarks. The defensive reasoning problem that matters for SOC work — noise, hypothesis stacking, long sessions — runs orthogonal to those benchmarks.
The Harder Finding: No Model Crosses 50%
Across 14 frontier models from Anthropic, OpenAI, Google, and open-weight providers, none crossed 50% coverage. Frontier LLMs routinely score above 80% on offensive security benchmarks. Defense is the structurally harder problem, and the gap between the two is growing wider as labs train for offense.
The practical upshot: raw LLMs should not be placed directly in front of a SOC queue. The benchmark tested the same scenarios with Simbian’s own harness wrapped around the best bare LLM — coverage went from 46% to 95%, independently verified by a global MSSP in April 2026. Forty-nine points came from context, skills, and agent loop design. One point came from the model choice.
What This Means for Anthropic’s Cyber Push
Anthropic has built real infrastructure around cybersecurity: Glasswing (200+ organizational partners, 10,000+ vulnerabilities patched), Claude Mythos (NSA access, classified deployments), and the Opus 4.7 Cyber Verification Program. The lab’s positioning in security is genuine. But the Simbian numbers surface a gap between offensive safeguard quality and defensive capability quality that marketing has not bridged.
The recommendation from the benchmark data is to run Opus 4.6 for SOC investigation work and wrap it in an agentic harness. Sonnet 4.6 is the price-performance pick at $2.21 per investigation. Skip Opus 4.7 for this specific workload.