GLM-5.2 Released with 1M Context and MIT License

Z.ai (formerly Zhipu AI) has released its latest flagship model, GLM-5.2. Its standout feature is a **1 million-token context window** that operates reliably. Compared with the previous-generation GLM-5.1, it delivers significantly improved performance on long-running tasks.

On the architecture side, it introduces a new mechanism called "IndexShare," which reuses the same indexer across 4 sparse attention layers, cutting per-token FLOPs at a 1M-token context length by 2.9x. It also improves the MTP layers, extending the acceptance length for speculative decoding by up to 20%. For coding, the model adopts a flexible design that lets it choose from multiple thinking-depth levels.

On benchmarks, it scores 99.2 on AIME 2026 and 91.2 on GPQA-Diamond. It puts up 62.1 on SWE-bench Pro, demonstrating frontier-class performance, and scores 76.8 on the agentic benchmark MCP-Atlas.

The model is available as fully open source under an **MIT** license with no regional restrictions. An FP8-quantized version is already published on Hugging Face and has hit the trending charts.

New Agentic Benchmark "Briefcase": Claude Fable and GLM-5.2 Among the Top Performers

Artificial Analysis has released a new agentic benchmark called "Briefcase." The benchmark tests an LLM's ability to plan and execute tasks, and since existing benchmarks have become saturated, it is said to enable more practical evaluation.

In the initial results, Anthropic's Claude Fable and GLM-5.2 each post top-class scores in their respective cohorts, underscoring how competitive the latest models are in agentic capabilities.

A Sparse Attention Breakthrough: Subquadratic Proves What "SubQ" Can Do

MIT Technology Review has published an in-depth report on the Miami-based startup Subquadratic. The company has developed a new LLM called "SubQ" that abandons the "dense attention" at the core of the Transformer in favor of **sparse attention**.

In conventional Transformers, the "quadratic blowup"—where doubling the text length quadruples the compute—has long been a problem. Instead of comparing every token with every other token, SubQ dynamically extracts only the important relationships, drastically reducing computational cost. The result is processing **56x faster** than models using FlashAttention. It records 89.7% on LiveCodeBench, on par with top-tier coding models.

In an independent evaluation by the third-party firm Appen, it achieved 98% retrieval accuracy on massive context windows of 6 to 12 million tokens. Running Anthropic's Opus 4.6 on the RULER 128 test reportedly costs $2,600, compared with just $8 for SubQ.

That said, skeptical voices remain, pointing out that SubQ is bootstrapped from reused Qwen weights and has not yet been widely released. Some are asking whether it is "the biggest breakthrough since the Transformer, or AI's Theranos," and how things unfold from here will be closely watched.

OpenAI's "Trait-Based Training": Making Models Safer Across the Board with a Small Amount of RL

OpenAI researchers have shown that when desirable behavioral traits such as truthfulness and corrigibility are trained with reinforcement learning, safety improves across multiple domains.

Interestingly, **a model trained on healthcare data also became better at detecting scam tactics**. Scores improved on 44 of 53 benchmarks, confirming a cross-domain generalization effect. The approach is drawing attention as a different route to safety from Anthropic's Constitution-based method.

Google Appeals German Court Ruling Over Liability for AI Search Results

Google has appealed a ruling by the Munich Regional Court in Germany. The court found that Google itself is directly liable after its AI search feature (AI Overviews) falsely linked two publishers based in Munich to a fraud case.

Google calls it a "minor error," but the court took a different view. If the ruling becomes final, it could set an important precedent for the scope of platform liability for AI-generated content, and it is expected to influence how AI search services operate across the EU.

Estonia to Grant Digital IDs to AI Agents

Estonia is planning to introduce a digital ID system for AI agents. Already known for e-Residency, the country is pursuing a pioneering initiative to give AI agents a legal and technical identity. As a policy design for an era in which AI agents autonomously carry out transactions and use services, it could become a reference case for other countries.

---