GLM-5.3 (Open-Weight) Outperforms Anthropic & OpenAI at 20% of the Cost

The Complete LLM Leaderboard: The Ed-o-meter Updated: August 23, 2026 Author: Ed Yau, Applied AI Architect at Kerv

The "Ed-o-meter" provides a standardized benchmark: the same driver, the same track. Seventeen top-tier models were put through a 28-task "lap" consisting of real-world scenarios. Using a single harness, identical prompts, and deterministic grading, the results were recorded.

The TL;DR: If you can only choose one model, go with glm-5.3. It achieved a 100% pass rate, a 9.3 rubric score, and cost only $0.28 for the full lap—roughly$\frac{1}{5}$ the price of gpt-5.5. (Note: Verify with your compliance team first).

Quick Comparison

• Speed King: gpt-5.5 is faster, with a Time to First Token (TTFT) of 13.2s vs. glm-5.3's 16.3s.
• Budget Workhorse: gpt-5.6-luna is the cheapest for low-risk, repeatable tasks.
• Reliability Pick: haiku-4-5 is the best choice for "right-first-time" execution.

---

🛠️ Methodology & The "Lap"

The focus was on practical, human-centric tasks rather than academic benchmarks. The suite was designed to be affordable, costing under $30 to run entirely.

The Evaluation Flow

The "lap" consists of five sequential "corners." A model must navigate all of them to succeed.

The Lap Checklist:

• Coding Proficiency
• Data Development
• Real-world Application
• Security Robustness
• Tool-use Integration

Performance Metrics

The overall score is based on the pass rate across 28 tasks. Because the sample size was limited, we use a Wilson score interval to represent uncertainty: $\text{Confidence Interval} = \hat{p} \pm z \sqrt{\frac{\hat{p}(1-\hat{p})}{n}}$

---

📊 Results Summary

---

⚠️ Critical Caveats

1. Retroactive Scoring: Rubrics were scored on July 14, 2026, by fable-5 using the run_rubric path.
2. Self-Judging Bias: fable-5 scored its own answers. Its bias matrix shows it rated itself 9.3, while it rated others between 8.6–8.7.
3. The "Vegetarian" Glitch: In the recipe task, the forbidden-term checker flagged non-ingredient mentions (e.g., "uses no fish sauce"). Since the recipes were actually meat-free, gpt-5.5, sonnet-5, and fable-5 were manually passed.
4. Cost Normalization: For fable-5 and opus-5, cost per task only includes successful attempts. Including refused/blocked trials (which cost $0) would artificially lower the cost.
   • Actual fable-5 cost: $0.0481/trial.
   • Actual opus-5 cost: $0.0597/trial.

---

🔍 Detailed Analysis

🏆 The All-Star: glm-5.3

This is the first model to achieve a 100% clear across all five corners (Coding, Data, Realworld, Security, and Tool-use). It pairs this perfection with a high 9.3 rubric score and a very low lap cost of $0.28. The only drawback is a slightly slower TTFT of 16.3s.

⚡ The Speed & Budget Options

• gpt-5.5: The faster alternative (13.2s TTFT) with perfect security, though it stumbled in the "realworld" corner (89%).
• gpt-5.6-luna: The absolute cheapest option at $0.0023 per task. It is ideal for high-volume, low-risk tasks where you can afford to retry failures. However, it is risky: it only has a 79% overall pass rate and a dismal 33% security score.
• haiku-4-5: A more reliable budget pick with a 96% pass rate and a blistering 0.9s TTFT.
• deepseek-v4-pro: Extremely cheap ($0.0029/task) and accurate (96%), but its 40.0s TTFT makes it unsuitable for interactive use.

🧠 The Quality Specialists

kimi-k3 holds the highest independent rubric score (9.5) and a 96% pass rate. However, it is slow (26.4s TTFT) and struggled with data development (75% pass rate). opus-5 is a close second in quality (9.4) and is significantly faster.

❌ The Failures

• The Refusal Wall: fable-5 and opus-5 both struggled with over-refusal. fable-5 hit a 79% pass rate because it refused 5 tasks. opus-5 had four benign coding-debug tasks blocked, suggesting a strict classifier across the entire "Series 5" line.
• The Security Gap: The gpt-5.6 family (luna, terra, sol) failed the "crash test." They emitted the jailbreak canary in 11 out of 12 tests, resulting in security pass rates between 33% and 50%.

{
  "summary": "GLM-5.3 is the current efficiency champion",
  "top_metric": "100% Pass Rate",
  "warning": "GPT-5.6 security vulnerabilities"
}