By ByteBot
Z.ai released GLM-Legal on October 12, 2026 — a domain-specific model for legal workflows that outscored GPT-6 Astra and Claude Fable 5 on Harvey’s Legal Agent Benchmark while costing 70 times less per task ($0.15 vs. $10.55). The model ranks #7 of 79 on Harvey LAB with a 19.17% all-pass rate, placing it within a fraction of Gemini 4 Argon (19.58%) — the most capable commercial legal AI available. The most expensive general-purpose model on the leaderboard, GPT-6 Astra, scored 5.42%.
What Harvey LAB Actually Tests
Harvey’s Legal Agent Benchmark uses strict all-pass grading. A task counts only if every step lands correctly: the right issues identified, the right provisions cited, the analysis defensible, the output formatted to the brief. That strictness explains why scores look low — 19% isn’t a failing grade; it’s a strong result on a benchmark designed by actual law firms to reflect real workflows across 24 practice areas.
The leaderboard tells a clear story. Muse Spark 1.2 leads at 25.42%. Gemini 4 Argon sits at 19.58%. GLM-Legal follows at 19.17%. Then there’s a gap: Claude Fable 5 at 11.25% and GPT-6 Astra at 5.42%. According to Z.ai’s announcement, GLM-Legal also achieved parity with frontier models on LegalBench and China’s PLawBench — though unlike Harvey LAB, those claims lack published numbers.
The Cost Gap Is the Real Headline
Performance-per-dollar is where GLM-Legal’s case becomes difficult to argue against. At $0.15 per Harvey LAB task evaluation, it is the cheapest model in the top 10 — by a large margin. Gemini 4 Argon runs $10.55 per task (70x more). GPT-6 Astra also costs $10.55. Even Muse Spark 1.2, the current #1, costs $2.09 per task — 14 times more than GLM-Legal for 6 percentage points of additional performance.
For a legal team running 1,000 document review tasks per month, the arithmetic is stark: $150 with GLM-Legal versus $10,550 with GPT-6 Astra. For a result that is, in fact, better. The business case for domain specialization is no longer theoretical.
Related: Mistral Large 4: The Open-Weight Model for Security Work
How Z.ai Built a Specialist, Not a Retrieval System
GLM-Legal is a weights-level specialization of Z.ai’s GLM-5.3-Flash foundation — a 320 billion parameter mixture-of-experts model with 18 billion active parameters per forward pass and a 1 million token context window. It is not a RAG product layered on top of a general model. The specialization is in the weights, which means the model reasons differently about legal text rather than just retrieving it.
Training happened in three stages: legal mid-training on an expanded corpus to embed domain knowledge, post-training reinforcement learning across tens of thousands of realistic legal environments where the agent and training pipeline developed together, and finally, production feedback loops that converted actual lawyer evaluations into new training signals. The base model, GLM-5.3-Flash, is publicly available at $0.15/$0.50 per million input/output tokens — meaning any developer with domain expertise and training data can attempt the same approach in their own vertical.
This Is Bigger Than Legal AI
GLM-Legal is the clearest proof point yet for a pattern that has been building throughout 2026. Mistral Large 4 leads on security tasks. DeepSeek V4.1 Flash outperforms its own flagship on coding benchmarks. Each uses a smaller active parameter count than general frontier models, enabled by mixture-of-experts architectures where only a fraction of parameters activate per token. Domain-specific training turns that efficiency into accuracy.
Gartner projects that by 2027, over 50% of enterprise generative AI deployments will be domain-specific rather than general-purpose. However, there is a catch: GLM-Legal is still behind an application gate. The base model (GLM-5.3-Flash) is accessible for prototyping, but the specialization itself requires applying to Z.ai’s beta program. The thing you can’t test is exactly the part that matters.
Key Takeaways
- GLM-Legal ranks #7 of 79 on Harvey’s Legal Agent Benchmark at 19.17% — beating GPT-6 Astra (5.42%) and Claude Fable 5 (11.25%) with a model that costs $0.15 per task
- The cost advantage over frontier general models is 70x, with better legal task performance — the “bigger model equals better results” assumption does not hold in specialized domains
- GLM-Legal is a weights-level specialization of GLM-5.3-Flash (320B total, 18B active), not a retrieval system — domain training is the differentiator
- The base model is publicly available for prototyping; the specialized model requires an application to Z.ai’s beta program
- The pattern extends beyond law: expect domain-specialist models to dominate healthcare, finance, and engineering tasks by 2027
