Reuters reported that Alibaba launched the new Qwen3.5 model on February 16, targeting 'agentic AI' task execution, emphasizing its ability to understand multi-step instructions and autonomously plan, make decisions, and invoke tools. The report noted significant performance improvements over the previous generation, enabling multi-round task decomposition with less human supervision. Official use cases include automated office workflows, intelligent customer service orchestration, and code generation. This release is seen as a major iteration in Alibaba's large model product line, shifting focus from dialogue to executable workflows and laying foundational capabilities for future agent applications.

TrendForce reported that Zhipu AI increased prices for GLM-5's GLMCodingPlan in February 2026, effective February 12: domestic subscription prices rose by at least 30%, while existing users can retain their original rates; overseas subscriptions increased by approximately 30%–60%, and API usage fees were raised by about 67%–100%. The report suggests this reflects a shift in China's LLM market from subsidy-driven discounting toward tiered pricing and profitability, driven by rising demand and computing costs. The article mentioned that Moonshot AI had already raised certain API prices in 2025, while Alibaba, Tencent, and Baidu are promoting tiered billing models.

Fortune reported that Peter Steinberg, developer of the open-source AI agent OpenClaw, announced he has joined OpenAI to work on next-generation personal AI agents. OpenAI CEO Sam Altman stated that OpenClaw will remain open source and be transferred to a newly established independent foundation, sponsored by OpenAI with model and computing resources. The article recalled that the agent can handle tasks like email organization, meal ordering, and check-ins, integrating with messaging apps; however, it also highlighted security risks due to access to private data and external content. There were prior reports of the agent失控 mass-messaging after being connected to iMessage.

A subscription newsletter reported that OpenAI introduced 'Lockdown Mode' for ChatGPT, offering stricter usage restrictions and security warnings for high-risk workflows. Features prone to prompt injection attacks are now labeled as 'high risk.' This setting aims to reduce the likelihood of agents being manipulated into leaking data or performing unauthorized actions when browsing web pages, reading external files, or invoking tools. At the same time, OpenAI discussed faster inference modes and specialized chip acceleration but emphasized that security protections must align with deployment scenarios. For enterprise users, this mode helps mitigate misuse risks in shared accounts, outsourced collaboration, or sensitive data processing.

PYMNTS reported that uncertainty around AI investment returns triggered a tech stock pullback, wiping over $1 trillion from the combined market value of Amazon, Google, Meta, and Microsoft. The report indicated investors are increasingly scrutinizing the alignment between capital expenditures and revenue, with these companies expected to spend over $600 billion on capex this year. Citing research, the article noted that about one-third of CFOs believe agents will significantly impact cash holdings and payment timing decisions, while 35% see them as useful for standardizing accounting language and internal transaction reporting.

TLDR News reported that Stripe acquired billing platform Metronome for $1 billion, as its existing system struggles to meet the real-time, high-concurrency, and usage-based billing needs common in AI products. Metronome excels at converting fine-grained usage such as API calls into invoices, supporting more flexible pricing and settlement. The acquisition is seen as Stripe strengthening commercial infrastructure for the AI era, particularly for SaaS and agent invocation scenarios, allowing customers to control costs while rapidly iterating billing strategies. Such M&A activity reflects a broader shift in AI company revenue models from subscriptions to usage-based and real-time settlements.

An email digest reported that Moonshot AI launched Kimi Claw, embedding the OpenClaw agent framework directly into the browser, enabling users to trigger skills and automation workflows within web pages. The product claims support for over 5,000 community skills and offers 40GB of cloud storage; it also supports 'bring your own Claw' to connect third-party OpenClaw deployments and reuse existing configurations. This move indicates the growing integration of OpenClaw's ecosystem into higher-level applications, with browsers becoming entry points for agent execution. For individual users, such packaging bundles skill markets, permission management, and cloud storage within a single interface.

Superhuman's newsletter reported that AI search company Exa launched 'Exa Instant,' optimized for low-latency retrieval in AI agent workflows, claiming response times under 200 milliseconds. This design suits frequent external information queries in multi-step tasks, improving efficiency in RAG, web scraping, and fact verification, reducing search as a performance bottleneck. The newsletter also suggested that such 'search layer' infrastructure is evolving from human-facing search experiences to API-first, programmatic forms. For automation scenarios requiring second-level decisions—such as customer service, trading, and monitoring—this low-latency retrieval enhances overall throughput and stability.