Anthropic Reduces Claude Distillation by 99%, Meta Releases Frontier Agent Models

What Actually Changed

Anthropic announced it has reduced the extraction of Claude's internal reasoning tokens by 90 to 99 percent, attributing large-scale distillation attacks to Chinese companies Zhipu and Moonshot, and has banned tens of thousands of accounts. Meta released Muse Wonder and Muse Wonder Full, described as its most capable models to date, built for personal and professional agentic tasks. Z.ai's GLM-Legal, built on GLM-5.3-Flash, scored 19.2% all-pass on Harvey's Legal Agent Benchmark, nearly 3x its base model performance.

Why It Matters

The significant reduction in distillation attacks represents a major defensive milestone for Anthropic, addressing illicit model copying that undermines the development proportionally robust safeguards. Meta's focus on agent-capable models signals a strategic shift toward task-specific AI systems rather than general-purpose models. The specialized performance of GLM-Legal demonstrates the growing value of domain-specific models over general ones, particularly in professional applications where accuracy and reliability are paramount.

Capability and Benchmarks

Z.ai's GLM-Legal achieved 19.2% all-pass on Harvey's Legal Agent Benchmark, significantly outperforming Claude Fable 5 and Kimi K3. The model is priced at $0.15 per test and requires application-based access. Anthropic's Dynamic Workflows for Claude Agents were detailed, drawing on Unix and Linux principles to help agents tackle larger projects without overwhelming developers. Research on Elastic Latent Recurrence (ELR) presents a method for training looped language models without accuracy loss at one loop, offering potential efficiency gains.

Compute, Cost and Pricing

GLM-Legal is priced at $0.15 per test, suggesting a premium pricing model for specialized legal AI applications. Anthropic's distillation defense measures likely require significant computational resources, potentially impacting operational costs. The research on Elastic Latent Recurrence (ELR) presents opportunities for more efficient model deployment, where inference compute can be traded for accuracy by changing loop counts without substantial performance penalties at baseline.

What We Are Watching

We're monitoring the effectiveness of Anthropic's distillation defenses over time and whether they will become industry standard. The emergence of MSL as a frontier lab with Claude 5.5-tier capabilities suggests potential new competition in the premium model space. Anthropic's warning that small AI models may be capable of self-replication across cheap IoT devices without datacenter access represents a significant security consideration for the industry's future development.

Where Consensus Is Wrong

There's a growing assumption that larger general-purpose models will dominate, but the superior performance of specialized models like GLM-Legal suggests the future may belong to domain-specific architectures. The industry focus on general intelligence may be misplaced when specialized models demonstrate superior performance in critical applications. The belief that model theft is primarily an issue for large companies may be incorrect, as evidenced by the industrial-scale distillation attacks targeting frontier models.