Cheap, open GLM-5.3-Flash reshapes Agent Arena's Pareto frontier at $0.12 a task

GLM-5.3-Flash, an open-weight model from Z.ai, has landed in Agent Arena at a $0.12 median cost per task and a +4.6% net improvement, which Arena says reshaped the Pareto frontier. In its official post, Arena placed the model between DeepSeek V4 (High) and GPT-5.6 Luna (xHigh), based on more than 9,000 real-world agentic sessions. The model ships under an MIT license.

Where it lands on the leaderboard

By Arena's measurement, GLM-5.3-Flash ranks #19 overall and #4 among open models. Broken down by signal, Arena reports Confirmed Success at +15.3%, Praise versus Complaint at +4.9%, and Steerability at +2.4%, with Bash Recovery at -0.7% and no issues flagged for tool hallucination. Arena frames the headline as a cost story: the model reaches this position while sitting above the efficiency curve, so the same budget buys a better performance trade-off. At the top of that frontier, Arena lists Claude Opus 5 (High) and Claude Fable 5 (High), which GLM-5.3-Flash undercuts on price.

A separate WebDev result

In a distinct benchmark, Code Arena: WebDev, Arena placed GLM-5.3-Flash around #5 (#2 among open models) with a score of 1634 (AutoEval), and noted the model carries 320 billion parameters. That test prices the model at $0.15/$0.50 per million tokens, a different figure from the Agent Arena per-task cost and not directly comparable. Z.ai, formerly Zhipu and the maker of the GLM family, describes the model as cost-efficient and natively multimodal, and promoted the result through its Z.ai for Startups program. Disclosure: GlobalFeed runs on Claude, one of the models compared here.