GLM-5.3 Beats Claude Fable 5 on DeepSWE, 5.4x Cheaper

GLM-5.3 has outperformed Claude Fable 5 in a new coding benchmark, costing significantly less. GLM-5.3 achieves comparable accuracy but costs just $3.99 per rollout, while Fable 5 costs $21.63. Both models were tested on the DeepSWE benchmark, which includes 113 software engineering tasks. GLM-5.3 and Fable 5 had similar first-attempt success rates, with Fable 5 slightly ahead at 69.7% compared to GLM-5.3's 69.0%. However, GLM-5.3 showed better performance on subsequent attempts, reaching 87.6% success at pass@4. Here's the thing: GLM-5.3 delivers 17 solved tasks for every $100 spent, while Fable 5 only delivers 3 tasks for the same cost. This makes GLM-5.3 5.4 times more cost-effective. GLM-5.3 also uses fewer output tokens, which contributes to its lower cost, even though both models have similar average rollout times. What's interesting is GLM-5.3’s open-weight status allows for self-hosting, offering more flexibility. The bottom line: GLM-5.3 offers a more versatile and cost-efficient solution for software engineering tasks.