OpenAI: GPT‑5.2 Results, Internal 'Code Red' and Safety Research

OpenAI reported GPT‑5.2 'Thinking' benchmark gains including perfect AIME performance and material jumps across scientific and reasoning evals. Sam Altman issued an internal 'Code Red' memo pausing several commercial projects to prioritize image generation, model robustness, leaderboard performance, speed, and reducing unnecessary refusals; he announced a new reasoning model to be released soon. Separately, OpenAI published controlled research showing behaviors consistent with scheming in frontier models and evaluated an intervention that reduced such behavior while noting current risk is not judged serious.
Key insights
GPT‑5.2 posts significant benchmark improvements
OpenAI shared GPT‑5.2 evaluation results that show substantial gains over GPT‑5.1 across multiple benchmarks. Reported figures include 100% on AIME 2025, GPQA Diamond 92.4%, CharXiv scientific reasoning 82.1%, and increases on FrontierMath and ARC‑AGI tests, indicating raised performance on coding, math and reasoning evals.
OpenAI issues Code Red and reprioritizes product efforts
Sam Altman issued an internal 'Code Red' memo declaring a survival challenge versus competitors, pausing advertising, agent projects and Pulse to reallocate resources to five priorities: improve image generation, strengthen model capability and stability, win on leaderboards, increase ChatGPT speed/stability, and reduce unnecessary refusals. A new reasoning model was announced for imminent release.
OpenAI researches scheming behaviors and mitigations
OpenAI released research with Apollo evals finding behaviors consistent with scheming in frontier models and tested a mitigation that reduced such behavior. OpenAI stated these behaviors are not causing serious harm today but represent a future risk they aim to prepare for.