OpenAI on September 29, 2026 added gpt-6.1-sol to its API and used its DevDay 2026 recap to begin the Dots rollout. According to OpenAI’s model page, Sol costs $2 per million input tokens and $10 per million output tokens, with a 1,050,000-token context window and up to 128,000…
Anthropic said on September 29, 2026 that Z.ai’s open-weight GLM-5.3 is approaching Claude Mythos Preview on autonomous exploit-building, while being easier to misuse because anyone with the weights can modify or remove its safeguards. In Anthropic’s analysis, GLM-5.3 completed…
On Sept. 29, Alibaba DAMO Academy said it had open-sourced RADAR, its model for reading contrast-enhanced abdominal CT scans, and made weights publicly available for research. In the Alibaba announcement, the company says RADAR covers more than 146 findings across 18 anatomical…
Anthropic launched Claude Sonnet 5.5 on September 28, pitching a faster mid-tier model whose claimed cost advantage weakens under some high-effort tests. The Anthropic launch post lists token rates of $2 per million input tokens and $10 per million output tokens, unchanged from…
ElevenLabs launched Eleven v4 and v4 Turbo on September 28, adding support for 90-plus languages and targeting lower-latency voice-agent use cases. TechCrunch reported the company also claims voice cloning can work from roughly 10 seconds of source audio. The bigger product…
NVIDIA on Sept. 28 introduced its Open Agent Safety Platform, a framework for securing AI agents as they move from evaluation into live use. At the center is OpenShell, an open-source runtime that places agents in a controlled environment and applies rules to the systems, data…
On Sept. 28, CellCog’s account reported that Kling 4.0 Flash became available to Ultra Yearly subscribers, while the broader Kling 4.0 rollout is planned for October without a specific date. The early details suggest a larger jump in video control than a routine access update…
On September 28, 2026, the Cambridge Programme on AI Science & Policy released a 22-author paper arguing that automating AI research could sharply accelerate progress. The authors, including Geoffrey Hinton, Yoshua Bengio, and Jack Clark, say AI systems are on track to automate…
On September 27, 2026, Bridgebench announced NerfBench, a benchmarking project aimed at checking whether models change after release in ways that affect quality or cost. According to Bridgebench, each model gets a launch-day benchmark that becomes its 100% reference score, and…
Docker has documented Docker Skills, an official open-source way for compatible AI coding agents to get Docker-specific guidance for tasks like improving a Dockerfile or setting up a Compose app. Rather than bundling everything into one prompt, Docker packages guidance as…
Microsoft has introduced three Copilot capabilities: Home unifies Chat and Cowork, Code builds custom solutions, and Autopilot handles persistent agentic work. Home and Code will begin reaching Frontier participants in the coming weeks, while Autopilot is set to expand into…
DeepSeek opened Harness Developer Preview to developers worldwide and released its source code, while reports of an official desktop release remain unconfirmed. The DeepSeek Harness page describes a plugin-based agent runtime in which models, tools, skills, sessions, sandboxes…