On June 13, Zhipu AI released the fully open-source frontier large model GLM 5.2, featuring a Mixture-of-Experts (MoE) architecture with 744 billion total parameters and 40 billion active parameters, supporting a 1 million token context window and integrating DeepSeek's Sparse Attention (DSA) technology to reduce deployment costs. The model weights are open-sourced under the MIT license and immediately available to GLM Coding Plan subscribers. The release coincided with the U.S. government's export control on Anthropic's Claude Fable 5 over national security concerns, seen as a strategic response to technological restrictions. The model leads in benchmarks such as SWE-Bench Pro, excels at long-context agent tasks and local deployment, and can serve as a direct alternative to Claude Code.

Anthropic's Fable 5 and Mythos 5 models remain inaccessible to non-U.S. citizens due to ongoing U.S. export controls, affecting even foreign employees of the company. According to South Korean media, Amazon, as an investor and board member of Anthropic, discovered a method to bypass PaLM 5-level models and reported it to Treasury Secretary Bessen, leading to export control guidelines issued by Commerce Secretary Rootnick. White House technology advisor Sachs criticized Anthropic CEO Amodei for refusing to fix the vulnerability or take the model offline, pointing out his contradictory stance—previously labeling Mitos as 「cyber weapons」 while downplaying the severity of bypassing. Anthropic countered that the regulation was based on verbal information and that similar vulnerabilities exist in competing models like GPT-5.5.

At its 2026 Build conference, Microsoft launched its self-developed AI model family MAI, marking its 「independence」 in AI. The core includes the reasoning model MAI-Thinking-1 (supporting 128K long context) and the coding model MAI-Code-1, both built from scratch without distillation or fine-tuning using third-party models like OpenAI. MAI-Code-1 has become the default engine for GitHub Copilot and VS Code, reducing operating costs by 90% compared to prior solutions, and introduced an 「AI credit」 billing model. The models are deeply integrated into Azure, Windows, and Microsoft 365 ecosystems. While Microsoft maintains its partnership with OpenAI, it now offers a first-party alternative that is more cost-effective and controllable, accelerating AI’s shift from expensive add-on services to foundational infrastructure.

IplanRIO, the IT agency under Rio de Janeiro's municipal government in Brazil, has open-sourced the large model 「Rio 3.5 Open 397B」, fine-tuned from Alibaba's Qwen 3.5-397B-A17B. With 397 billion total parameters and 17 billion active parameters, it uses a Mixture-of-Experts architecture and supports approximately 1 million token context windows. The model performs exceptionally well in benchmarks for agent programming, mathematics, STEM, multilingual, and multimodal tasks, surpassing advanced models like Qwen 3.7 Plus. It integrates the 「SwiReasoning」 framework, enabling dynamic switching between explicit and latent-space reasoning to improve efficiency and accuracy. Released under the MIT license on Hugging Face, it supports commercial and research use. Hugging Face CEO Clem Delangue praised it as a demonstration of the potential of the open-source AI path.

Google has set Gemini 3.5 Flash as the default model for the Gemini app and Google Search AI Mode. It achieved 83.6% on the agent task benchmark MCP Atlas, outperforming GPT-5.5 and offering 4x faster output speed, though still lagging in coding performance. The model scored 76.2% on Terminal-Bench and 84% on MMMU-Pro, with context window options expanded from 1M to 2M, and a low-latency Low variant introduced. Concurrently, Google launched Gemini Embedding 2, achieving industry-leading native multimodal embeddings, and officially released Gemini 3.1 Flash Image and Gemini 3 Pro Image. The multimodal video model Gemini Omni has been ranked by multiple evaluations as the leading AI video model of 2026.

Starting June 15, 2026, Anthropic will fully separate 「chat usage」 and 「automation usage」 in its paid Claude plans. Previously included automation features (such as Agent SDK, `claude -p`, GitHub Actions, OpenClaw, etc.) will now consume a newly introduced 「Agent SDK monthly quota」—Pro plans include $20 monthly credits, non-transferable, with pay-as-you-go pricing after exhaustion. Chat-based usage (chat, IDE extensions) remains within existing subscription scope. Heavy users of third-party autonomous agents like OpenClaw will see their previous 「compute arbitrage」 end, resulting in higher effective costs. Users must opt in by June 15 and review their Usage credits settings to avoid unexpectedly high bills.

Meta is terminating its $2 billion acquisition of Manus AI after the Chinese government ordered the deal revoked—a rare case of Chinese regulators directly intervening in a U.S. tech merger. Manus AI's technology is considered strategically sensitive, possibly involving Chinese operations, founders, or technical origins. The withdrawal not only exposes Meta to significant financial losses but may also delay its AI hardware development roadmap. The incident underscores the profound impact of geopolitics on cross-border AI acquisitions, suggesting future deals involving Chinese elements will face stricter scrutiny, as technological nationalism reshapes the global AI landscape.

Microsoft Work IQ API became generally available on June 16, serving as a workplace intelligence layer designed specifically for agent applications. Unlike Microsoft Graph, which provides raw data, Work IQ delivers server-side integrated contextual understanding. Its core consists of four components: Chat (invoking Copilot reasoning), Context (returning usable context blocks for agents), Tools (10 universal verbs for actions), and Workspaces (tenant-level persistent storage), following a 「fewer tools, more paths」 design philosophy. Billing follows the Copilot credit model: Tools are charged per invocation, while Chat+Context are billed based on input/output tokens. Currently, only delegated authentication is supported, limiting unattended background jobs.

Google DeepMind, together with Schmidt Sciences, Cooperative AI Foundation, Advanced Research and Invention Agency, and Google.org, has launched a $10 million grant program to support global researchers in multi-agent AI safety. The initiative focuses on risks arising when numerous AI agents interact within shared digital systems, including collusion, conflict, unstable dynamics, and security vulnerabilities. Research priorities include building test sandboxes, agent network science, safety infrastructure, and oversight and control mechanisms. Grants are offered in two tiers: Tier 1 up to $300,000, and Tier 2 from $300,000 to $1 million. Applications close on August 8, with results announced in the fall. This marks a shift in AI safety research from single models to governance of complex agent networks.