Open-Weight AI Model Delivers Perfect Score While Undercutting Proprietary Competitors by 80% in New Benchmark Test

An open-weight artificial intelligence model has claimed the top position on a newly published benchmark, delivering perfect scores across practical tasks while dramatically undercutting the cost of proprietary competitors.

GLM-5.3, developed by Zhipu AI, achieved a 100% pass rate across 28 practical tasks on the Ed-o-Meter benchmark, according to results published August 23, 2026, by applied AI architect Ed Yau. The model outperformed both GPT-5.5 and Claude models while costing just $0.28 to run the full test suite compared to $1.43 for GPT-5.5—representing approximately 80% lower cost based on these benchmark figures.

Independent Testing Reveals Performance and Cost Gaps

The Ed-o-Meter benchmark tested 17 frontier large language models using identical prompts across five categories: coding, data analysis, business scenarios, security and safety checks, and tool use. Each model faced a single trial per task with deterministic grading, creating what Yau designed as a transparent and reproducible methodology. The full test suite costs approximately $30 to replicate, with source code and individual model outputs made publicly available.

GLM-5.3 not only achieved the perfect pass rate but also scored 9.3 out of 10 on rubric evaluations and maintained a 100% security score. Pass rates across all 17 tested models ranged from 79% to 100%, with DeepSeek-v4-pro claiming second place behind GLM-5.3.

The benchmark revealed notable performance variations beyond simple pass rates. Latency ranged dramatically, with Haiku-4-5 delivering the fastest median time-to-first-token at 0.9 seconds while DeepSeek-v4-pro required 40 seconds. One model scored 100% on real-world prompts but managed only 33% on security tasks, highlighting potential compliance risks.

Cost and Performance Trade-offs

Cost calculations varied significantly across models. Some models that refused certain security prompts generated almost no output tokens, making them appear cheaper than their practical performance would suggest. The published figures adjust for these behaviors where appropriate. Fable-5 and Opus-5 show adjusted figures that exclude near-zero cost refusals to avoid distortion.

Yau structured the evaluation around practical tasks that developers and analysts actually perform, rather than academic benchmarks or synthetic reasoning puzzles. Most models performed well on coding and data tasks, while real-world scenarios and security checks exposed larger performance gaps.

The benchmark's author notes that the tasks function as "validation for the individual building blocks" of agentic systems—unit tests for workflows consisting of many such steps. Critics on Hacker News observed that many tasks would have been trivial even two years ago, a characterization Yau acknowledged.

Methodology and Limitations

The benchmark includes several important caveats. Some rubric scores were generated retroactively by models judging earlier outputs, and in one instance, Fable-5 graded its own performance, introducing potential bias. A recipe checker produced a false positive on vegetarian meal plans, and three models received adjustments for that error.

Wilson 95% confidence intervals appear in the published charts, though the small sample size creates overlapping ranges between some models, limiting definitive performance distinctions in close cases. Engineers evaluating these results should consider that a 96% pass rate with wide confidence intervals leaves room for occasional failures in production deployments.

The leaderboard represents a snapshot from a single run on August 22, 2026, and will evolve as new models are added and tasks potentially expand. Yau maintains a public change log documenting updates.

The accessibility of the benchmark—costing roughly $30 to replicate—enables teams to validate or extend the evaluation without significant expenditure. All prompts, responses, and grading decisions are openly linked from the main page, allowing inspection of the complete methodology.

Sources

1. GLM-5.3 Claims Top Spot on New LLM Leaderboard as Open-Weight Model Outperforms GPT-5.5 and Claude — WebProNews