GLM-5.3 matches Claude Fable 5 on DeepSWE at a fifth of the cost

Together AI reports that GLM-5.3, an open-weight model, achieves 69.0% pass@1 on the DeepSWE benchmark, statistically tied with Claude Fable 5's 69.7%. The cost difference is stark: GLM-5.3 runs at $3.99 per rollout, while Claude Fable 5 costs $21.63. Per $100 spent, GLM-5.3 solves 17 tasks, Claude Fable 5 solves 3. The total batch cost for GLM-5.3 was $1,806, versus $9,346 for Claude Fable 5.

Retry behavior and failure patterns

GLM-5.3's pass@2 is 81.1%, and pass@4 is 87.6%, both higher than Claude Fable 5's 77.1% and 84.1%. Coverage, the fraction of tasks with at least one successful attempt, is 87.6% for GLM-5.3 and 84.1% for Claude Fable 5. However, Claude Fable 5 shows higher reliability at 82.0% versus GLM-5.3's 78.8%, and solves more four-for-four tasks: 56 versus 48.

Failure anatomy is similar: GLM-5.3 has 61% near-misses, 16% big-misses, and 11% regressions; Claude Fable 5 has 57% near-misses, 18% big-misses, and 11% regressions. Both models regress the existing test suite in 11% of failures, far below the GPT-family's 20% rate.

Domain and language strengths

GLM-5.3 wins five of eight task domains: structured query, language internals, stateful reactivity, concurrency, and program analysis. Claude Fable 5 wins three: data modeling/serialization, build/ops, and protocol conformance. By language, GLM-5.3 leads in JavaScript (90 vs 75) and TypeScript (61 vs 57), while Claude Fable 5 leads in Rust (85 vs 70) and Python (70 vs 66).

Per-task correlation between the models is 0.65, the highest among measured pairings. Combined, they cover 106 of 113 DeepSWE tasks, or 93.8%. This suggests a routing strategy: running GLM-5.3 first, then escalating to Claude Fable 5 for unresolved tasks, yields 81.1% accuracy at $10.74 per task.

Operational differences

GLM-5.3 outputs 80k tokens per rollout and takes 124 steps, while Claude Fable 5 outputs 114k tokens and takes 85 steps. Average rollout time is nearly identical: 35 minutes for GLM-5.3, 34 minutes for Claude Fable 5. GLM-5.3's peak context window is 155k tokens, Claude Fable 5's is 205k. Infra errors differ: GLM-5.3 recorded 1, Claude Fable 5 recorded 16.

Together AI concludes that GLM-5.3's higher ceiling on retries makes it the cost-effective default for most software-engineering tasks. Claude Fable 5's premium is justified only for Rust-heavy or serialization-critical workloads. Since GLM-5.3 is open-weight and Claude Fable 5 is closed and only via Anthropic, the routing strategy also balances cost with access constraints.