Z.ai's GLM-5.2 Beats Every Open-Weight Model at a Sixth the Cost

Z.ai's GLM-5.2 just became the highest-scoring open-weight model on the Epoch Capabilities Index (ECI), Epoch AI's composite benchmark that aggregates scores from over 50 distinct evaluations into a single general-capability scale. GLM-5.2 lands at an estimated ECI of 152, surpassing every other open-weight model Epoch has evaluated. That's a meaningful number: on the ECI scale, Claude 3.5 Sonnet is anchored at 130 and GPT-5 at 150, so GLM-5.2 is sitting right at the frontier of what closed labs were shipping not long ago.

The catch? It remains behind models like Gemini 3 Pro, released over 7 months ago. The open-weight frontier is catching up fast, but it's still chasing a moving target.

What the ECI Actually Measures

The ECI isn't a single benchmark -- it's a statistical aggregation. It works by stitching together results from dozens of evaluations, inferring each benchmark's relative difficulty from the overlap of models that appear on multiple tests. Harder benchmarks carry more signal. The scale is linear, meaning a 10-point jump at the top is supposed to be as meaningful as one at the bottom. The ECI is designed to capture broad, underlying capability useful across many tasks, rather than performance on specific skills.

For context on where GLM-5.2 sits on the broader leaderboard: Claude Fable 5 achieved a new high score of 161 on the ECI, beating GPT-5.5 Pro by 1 point. GLM-5.2 at 152 is behind the very top closed models, but it's the first open-weight model to get this close.

753 Billion Parameters, One Million Tokens

Beijing-based AI company Z.ai, formerly known as Zhipu AI, released GLM-5.2 on June 13, 2026, and the model has quickly become the most capable open-weight large language model available. The headline specs:

• 744B total parameters with 40B active per token -- it's a Mixture-of-Experts (MoE) architecture, meaning the full 744B parameters exist but only a fraction are activated per forward pass, keeping inference costs manageable
• A 1,000,000-token context window (labeled glm-5.2[1m]