Zhipu releases GLM-5.2 Coding Plan, with API access and open source release coming next week, supporting 1M context window. Tests show it can handle backend debugging and mini-program development, but lacks multimodal capabilities.
World Labs, co-founded by Fei-Fei Li, releases its first three papers, studying how mature 2D generative models can lower the barrier to 3D content generation; Flex4DHuman improves PSNR by about 9.3 dB on DNA-Rendering.
Anthropic says it has globally disabled Claude Fable 5 and Mythos 5 under a U.S. government export control order, affecting all customers and foreign employees; other Claude models are currently unaffected.
AnySearch attracted over 100,000 developers and exceeded 4 million calls within one month of launch, aggregating professional data such as finance and legal sources to help agents directly access usable information.
After receiving a $40 million investment led by OpenAI, Opal Electronics plans to launch its first AI wearable audio hardware in 3 to 4 months, designed to connect different large models to help users handle tasks.
City University of Hong Kong and China University of Petroleum researchers used unsupervised AI to analyze over 74 million water molecules, proving A/B components exist in ambient-pressure water and mapping their transition pathways, with findings published in Nature Physics.
Town raises $55 million in Series A funding led by a16z; its personal AI assistant connects tools such as email and calendars, remembers user habits, and proactively handles tasks.
UC Berkeley releases Agents’ Last Exam to test AI Agents on real computer tasks; GPT 5.5 achieves the top pass rate of 24%, beating Claude Fable 5’s 22%, while most models score 0 on the hardest tier.
Codex lead says token consumption grew strongly over the past 48 hours; OpenAI says Codex has over 5 million weekly active users, with about 14 million npm downloads in the past 30 days, below Claude Code's about 46.3 million.
Silicon Valley AI circles introduce Loop Engineering, using automated systems to replace manual agent prompting, automatically discovering, assigning, checking tasks and recording status, with Claude Code and Codex already supporting related modules.
OpenAI and Anthropic announced billion-dollar enterprise AI joint ventures on May 4, focusing on FDE to connect AI with corporate workflows, systems and knowledge so AI can perform real work.
YouMind releases version 1.0, upgrading creation features for writing, image generation, audio-video, web pages, and learning, with over 2,000 Skills listed, plus Sprite, iOS app, and API support to help users complete creation workflows more smoothly.
Shanghai AI Laboratory and others propose NeuroFlow, accepted by CVPR 2026. On the NSD dataset, it uses 25% of the training parameters to enable bidirectional conversion between vision and brain signals, helping reconstruct images from brain signals or generate brain signals.
Moonshot AI releases and open-sources Kimi K2.7 Code programming model, with 1.1T parameters, 30% average token consumption reduction, improved over-thinking problem, 10%-31.5% performance gains on multiple benchmarks, already available on Kimi API platform, 5-6x faster high-speed version will launch next Monday.