Tags
“Temporary versus permanent” is the line Meta’s new Muse models learned to draw in what you tell them. On Meta’s own eval, they now use critical context in 90% of tasks, up from 76%. How many days back do you want an assistant to remember your chats?
Google cut Gemini Omni 1.2 Flash to $0.07 per second of 720p video, 30% below Omni 1.1, though fast sports action and complex physics still need work. How many seconds does your typical generated clip run?
OpenAIanalysis
On OpenAI’s RemoteExploitBench, GPT-6 Astra refused 100% of tasks, while GPT-6 Cyber completed 95%, missing only one exploit it couldn’t get working remotely. Has a model ever flatly refused a legit security task of yours?
Six seconds per image is what OpenAI quotes for GPT-Image-2.6 Flare, but only with quality set to low. Sunburst, its sibling, aims at detail and steadier edits instead. Which quality setting do your image models run on by default, low or high?
Open modelsanalysis
“Our most capable 3T-class model to date,” Moonshot AI says of Kimi K3.1 Pro, and the open weights are now on HuggingFace under the company’s own license. Would you run Kimi K3.1 Pro on your own hardware or stick to an API?
