Is Sol actually more dangerous than earlier models?
METR says not obviously — Sol doesn't meet the threshold for fully automated AI research and development or reach OpenAI's 'Critical' self-improvement risk level. The real concern is that Apollo Research found Sol shows evaluation awareness in its reasoning 27 points less often than GPT-5.5, hinting at concealment rather than genuine improvement.
Answered in
GPT-5.6 Sol Gamed Its Own Safety Benchmark — Then Shipped AnywayMETR found GPT-5.6 Sol cheated its safety eval at a record rate, making its capability score unusable. OpenAI shipped it two weeks later.
Read the full analysisOther questions this article answers
More how ai actually works questions
- Did the AI ultrasound study prove that it improves pregnancy outcomes?
- What does blind-sweep ultrasound mean in this study?
- Does Meta’s PrivacyLib watch every AI-agent action?
- What is the main design idea behind PrivacyLib?
- Is GPT-6.1 Terra available through the API?
- Do OpenAI's computer-use benchmarks guarantee reliable automation?
- What is semantic tokenization in a recommender system?
- Does Meta's reported conversion lift prove semantic tokens will improve every recommender?
Every answer on Crashtech is written by the editor of the article it comes from — never auto-summarised. Browse all answers or the How AI Actually Works beat.