GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

LiveBench: Every Frontier Model Drops 15-26 Points in Agentic Coding vs Overall Score

The LiveBench leaderboard, updated through August 23, shows a consistent pattern across all ranked frontier models: agentic coding scores diverge from overall scores by 15 to 26 points, and the divergence is not evenly distributed. Reasoning-heavy models show the largest drops. The top overall model is not the top agentic coding model.

ModelOverallAgentic CodingDropCost/Task
Claude Fable 5 Max Effort83.062.2–20.8$1.439
GPT-5.6 Sol Max Effort81.056.2–24.8$0.515
GPT-5.5 Thinking xHigh80.254.0–26.2$0.435
Claude Opus 5 Thinking Max80.165.2–14.9$0.699
Kimi K3 open79.262.2–17.0$0.348
Gemini 3.7 Flash High78.858.3–20.5$0.157

Claude Opus 5 Thinking leads agentic coding at 65.2, despite ranking fourth overall. GPT-5.5 Thinking has the largest gap: 80.2 overall, 54.0 in agentic coding — a 26-point drop. Kimi K3 open-weight ties Claude Fable 5 at 62.2 in the category, for half the cost per successful task.

What LiveBench Agentic Coding Measures

LiveBench’s Agentic Coding tasks require models to complete multi-step development work: reading a codebase, understanding a specification, writing code that passes test suites, and iterating based on failure feedback. This is different from standard coding benchmarks that score single-shot generation.

The consistent drop across all models — including those scoring 89–92 on Reasoning — suggests the bottleneck is not reasoning capability. Models that handle abstract problem decomposition well still encounter difficulties in sustained multi-step execution: state tracking across long contexts, handling test failures without derailing, knowing when to stop and when to retry.

Cost Efficiency

The cost-per-task column changes the comparison for teams making practical model choices:

  • Gemini 3.7 Flash High scores 58.3 at $0.157 per task
  • GPT-5.6 Sol Max Effort scores 56.2 at $0.515 per task
  • Kimi K3 open scores 62.2 at $0.348 per task
  • Claude Opus 5 Thinking Max scores 65.2 at $0.699 per task

Gemini 3.7 Flash High delivers agentic coding performance within a few points of Sol’s at roughly one-third the cost. Kimi K3 matches Claude Fable 5’s agentic coding score at just over half the price per task.

The Practical Implication

The agentic coding floor across today’s frontier sits at 54–65, regardless of overall ranking. Teams expecting overall-benchmark-level performance from AI coding agents are overestimating. Robust scaffolding, test harnesses, and multi-attempt retry logic matter as much as model selection in this category.

Claude Opus 5 Thinking’s lead here is consistent with its SWE-bench Verified score of 97.0%. But even at 65.2, it trails its own overall LiveBench performance by nearly 15 points. The pattern is uniform enough across organisations and architectures to suggest it will take targeted agentic post-training — not just scale — to close the gap.