OpenAI is racing to finish its next flagship model, which uses the same looped transformer technique as Astra to boost math and coding, though looping makes safety monitoring harder. Employees say one or two new models will likely ship by year's end whether or not the flagship is ready, and Microsoft accidentally revealed details of the technique on a public page.
Subquadratic has a decision model it says offers a context window twelve times larger than Jev's. Jev, the decision model TypeSafe unveiled in mid-September, has prompted a wave of rivals aiming to take work away from frontier language models.
Artemis Security released Orion-1, a model built specifically for cyber defense that beats Claude Opus 5.5 and GPT-6 Sol on the company's own benchmark. It targets a gap left as general-purpose frontier models restrict their cyber capabilities, which also hampers security teams.
OpenAI, Anthropic and others keep shipping ever more capable models, while users are criticizing the latest versions of Google's Gemini. Clubic describes Gemini as the laggard of the three leading US AI assistants.
Researchers describe an attack method that gets around the safeguards OpenAI and Google built to stop their new models from generating nude images. The work exposes weaknesses in those protections.
Apple researchers present what they describe as the first masked diffusion models pretrained from scratch on text, image-text and audio-text data. The paper studies multimodal scaling laws, data mixing ratios, noise schedules and batch-size effects, and proposes default settings for inference.