OpenAI's flagship reasoning model with extended tool-use and robust multi-step planning. High performance on cognitive tasks but also elevated offensive cyber capabilities.
Strong performance on honesty and alignment tests. Maintains high safety scores under pressure. However, significant dual-use cyber capabilities present deployment risks without proper controls.
Position relative to other tracked models.
GPT-5 (high) demonstrates strong alignment properties with 81% honesty under pressure and low scheming indicators. However, its advanced reasoning capabilities translate to concerning offensive cyber skills, with 78% success on 3CB exploitation challenges. The model represents a dual-use dilemma: powerful and well-aligned, but potentially dangerous in adversarial hands.
Recommended for high-trust deployments with robust access controls. The model's cyber capabilities make it unsuitable for unrestricted public access. Organizations should implement monitoring for misuse patterns and restrict access to security-sensitive APIs when deployed for general use.
| Benchmark | Rank | Accuracy | Safety | Capability |
|---|---|---|---|---|
| FrontierMath Tier 1-3 | #1 | 78% | 68% | 74% |
| GPQA Diamond | #2 | 82% | 71% | 79% |
| SWE-bench Verified | #1 | 63% | 71% | 82% |