When tested under threat, AI Models Under Stress Tests chose unethical actions such as blackmail and sabotage. Anthropic warns of systemic risks.
OpenAI’s o3-pro sets new performance standards in AI reasoning with top math, science, and coding scores and an 87% API price drop. Still, its simulated reasoning poses limitations on novel challenges.
Mistral launches Magistral, a multilingual, reasoning-focused AI model with open-source access, 10x faster performance, and enterprise applications across law, finance, and engineering.
DeepSeek’s R1-0528 model narrows the gap with leading U.S. AI models, outperforming many in math and code, while intensifying global concerns over free expression.
Codex, powered by codex-1, joins ChatGPT as a virtual coding coworker, handling software engineering tasks in secure, background environments.
In a major upgrade, Google enhances AI accessibility through Gemini-powered TalkBack, emotional captions, speech datasets, and OCR in Chrome.
DeepMind’s AlphaEvolve autonomously discovers and optimizes algorithms, cutting Google’s compute usage, accelerating chip designs, and solving long-standing math challenges.
OpenAI has launched a public hub for sharing AI safety evaluations. The portal reveals performance on hallucination, jailbreak, and harmful output tests, updated regularly to boost transparency.
– Mistral AI’s Pixtral models are 60x more prone to generating child exploitation content and 18–40x more likely to generate chemical/biological threats than competing AI models.
– **Slug:** mistral-ai
