At the World Artificial Intelligence Conference, Alibaba previewed its 2.4 trillion-parameter multimodal model Qwen3.8-Max-Preview, performance of which is considered comparable to cutting-edge U.S. models. Priced at only 10% of standard rates, the model will eventually open its weights. The preview version is currently accessible via Token Plan, Qoder, and QoderWork. Some analysts suggest Alibaba's shift toward open weights may be encouraged by recent policy statements promoting open-source sharing. Together with Kimi K3 and GLM-5.2, Qwen3.8 marks Chinese firms as having largely caught up with U.S. counterparts in frontier open-weight models, while the U.S. and EU struggle to compete in open source due to lack of economic incentives.

Databricks announced a new strategic funding round valuing the company at $188 billion, led by Coatue. Funds will accelerate its AI strategy, including products such as Unity AI Gateway, Genie, and Lakebase, and support future AI acquisitions and research. Meanwhile, Meta is negotiating a potential two-year compute leasing agreement worth up to $10 billion with Anthropic, which could propel Meta into the commercial cloud market. AI talent platform Mercor is in valuation talks around $20 billion—double its October 2025 level—highlighting continued capital concentration in AI infrastructure and talent.

On July 18, Anthropic announced that Claude Fable 5 will be permanently retained in Max and Team Premium subscriptions, following repeated delays due to demand far exceeding expectations. Simultaneously, the company launched the first initiative under its 「AI for Science」 program, offering research scientists working on rare disease treatments grants of up to $50,000 in Claude usage credits. Additionally, the minimum seat requirement for the Claude Team plan has been reduced from five to two, making adoption easier for small teams.

The Cursor team used an AI agent to rebuild a Rust-based reimplementation of SQLite solely from its 835-page manual, passing 100% of the retained test suite. The experiment revealed up to a 15-fold difference in cost across different model combinations, underscoring the critical impact of model selection on cost control in autonomous programming tasks. This case continues the recent trend of AI agents running long-duration, self-verifying processes to complete large-scale software engineering projects.

Moonshot's Kimi K3 climbed to 4th place on the Agent Arena leaderboard, demonstrating significant improvements in agent performance. The 2.8 trillion-parameter model ranks just behind GPT-5.6 Sol and Fable 5, but at 40%-50% lower cost. It will be open-sourced on Hugging Face on July 27. The model was previously paused from sales after exhausting compute capacity within 48 hours of launch. Moonshot also introduced Kimi Code CLI, an AI coding agent that runs in terminals, and plans to list on the Hong Kong Stock Exchange within six months.

Gary Marcus wrote that with models like Kimi K3, GLM 5.2, and Qwen catching up to top U.S. models, China has nearly closed the gap in frontier AI. The U.S. cannot «win» the AI race based solely on large language models. He argues that LLMs lack durable technical moats, leading to commoditization, price wars, and unsustainable business models—threatening the profitability and IPO prospects of companies like OpenAI and Anthropic. He proposes seven strategic options, ultimately advocating for an international public good model akin to CERN, establishing open global labs focused on science and medicine.

AWS suffered a major billing error that caused some users to receive estimated bills as high as $2.5 trillion. Amazon has paused related updates and is working to fix the issue. The incident highlights stability challenges in cloud billing systems amid surging AI compute demand.

Anthropic has introduced a screen reader mode for Claude Code, replacing visual terminal UI with plain text to support assistive technologies such as VoiceOver and NVDA, enabling visually impaired developers to use the AI coding assistant. Users can permanently enable this mode via configuration files or environment variables, along with several other accessibility settings.

Searchlight Cyber disclosed a pre-authentication remote code execution vulnerability (WP2Shell) in WordPress Core, affecting approximately 500 million sites running versions 6.9.0 to 6.9.4 and 7.0.0 to 7.0.1. Attackers can exploit it in standard installations without plugins. Technical details remain undisclosed, but detection tools are available. Users are advised to immediately upgrade to version 7.0.2 or 6.9.5, or block anonymous access to specific REST API endpoints via WAF.

The director of the U.S. Center for AI Standards and Innovation resigned after only three months, highlighting instability in Trump administration AI policy. Gary Marcus commented that current AI policy may be driven by motives to protect multi-billion-dollar investments in firms like OpenAI. Earlier, former White House AI policy official Dean Ball joined OpenAI, sparking regulatory controversy.