<cite index="43-2">On DeepSWE v1.1, GPT-5.6 Sol leads single-shot at 72.7% pass@1 versus GLM-5.3's 69.0%, but GLM-5.3 reverses the order on retries: 81.1% pass@2 (tying Sol's 81.0%) and 87.6% pass@4 (beating Sol's 85.8%).</cite> <cite index="43-2">The key advantage: cascading GLM-5.3 first then escalating to Sol for failures solves 85.9% of tasks at $6.61 per task—13 points better than Sol alone at 72.7%, yet 21% cheaper than Solo solo at $8.37.</cite>
<cite index="43-2">Sol wins on latency and reliability: 19 minutes and 61 steps per rollout versus GLM's 35 minutes and 124 steps.</cite> <cite index="43-2">Sol's reliability stands at 84.5% with 61 tasks solved four-for-four, while GLM-5.3 trades reliability for broader coverage at 87.6% pass@4 but lower per-attempt consistency (78.8%).</cite> <cite index="43-2">GLM-5.3's failure mode is safer: it breaks existing tests in 11% of failures versus Sol's 20%.</cite>
For production routing: <cite index="43-2">the two models diverge sharply (0.43 per-task correlation) yet together cover 106 of 113 tasks (93.8%), making them ideal complements in a cascade.</cite> Domain split is clean—Sol dominates data modeling, serialization, ops tooling, and protocol work; GLM-5.3 excels at query languages, runtime internals, and reactivity. <cite index="43-2">GLM-5.3's standout is JavaScript at 90% (15 points ahead of Sol); Sol answers with Python, Go, and TypeScript strength.</cite> The engineering lesson: optimal coding agent cost is not a single model but a routing strategy.