AI News

IBM and Confluent are offering four IBM Granite Time Series foundation models in Early Access on Confluent Cloud. Through Apache Flink SQL, enterprises can apply forecasting, anomaly detection, optimization, similarity search, classification, and gap-filling directly to streaming data. Confluent provides native model inference, state management, scaling, governance, security, and downstream delivery, reducing the need for a separate machine-learning stack. The portfolio includes PatchTST-FM, FlowState, TTM, and TSPulse, which target different data volumes, time resolutions, and operational requirements. Availability starts on Confluent Cloud for AWS, with Confluent Platform planned for on-premises and hybrid environments.

Anthropic has launched Claude Fable 5.1 and Claude Mythos 5.1, sharing the same underlying model but offering different access and safety controls. Fable 5.1 is broadly available, while Mythos 5.1 is restricted to approved organizations. The models improve agentic science, coding, knowledge work and long-running tool use. Prompt-cache read prices fell 75%, potentially lowering highly agentic workloads by up to 45%, though maximum-reasoning tasks may cost more due to longer outputs. Anthropic also presented applications in root-cause debugging, protein design, Venus mapping and GPU optimization, positioning the release around deployable high-value work rather than low-cost general chat.

QuestMobile reports that targeted AI regulation is shifting investment toward productivity agents. China’s AI productivity-office segment reached 102 million monthly active users in July 2026, while Tencent, ByteDance and Alibaba established an early three-way competitive structure through integrations with WeChat, DingTalk, Feishu and related ecosystems. The report says future differentiation will come from workflow execution, organizational integration and closed-loop task completion.

Whop built an AI agent inside Slack that connects public channels, warehouse data, code changes, ad accounts, sales-call transcripts, and project tracking. The system is intended to make fragmented operational context available through one shared intelligence layer. Jonah Saya says the post will cover how Whop built it and what the team learned, while the supplied text does not include architecture, evaluation, accuracy, access controls, or safety details.