Back to benchmarks
Coding Agent BenchmarkVerified 25 days ago

GLM-5.3 vs Frontier Coding Agents: Terminal-Bench 3.0 Showdown

GLM-5.3's post-training gains push Terminal-Bench 3.0 from 4.6 to 28.3 — but how does it compare to Fable 5 and GPT-5.6 Sol on independent benchmarks?

GLM-5.3, Claude Fable 5, GPT-5.6 SolMultiple2026-08-19
Winner:Claude Fable 5

Results dashboard

Overall score

GLM-5.3
1769
Claude Fable 5
1743
GPT-5.6 Sol
1730

Results table

AgentScoreTimeCostPass
GLM-5.31769 Pass
Claude Fable 51743 Pass
GPT-5.6 Sol1730 Pass

What it is

A comparison of the three leading coding agents as of August 2026, using the one independently scored benchmark available on GLM-5.3's launch day: GDPval-AA v2, scored by Artificial Analysis (not by any of the model vendors).

Why this benchmark matters

Most coding agent benchmark numbers are vendor-reported. GLM-5.3's internal Code Bench and in-harness cyber scores are all Z.ai-run. Claude Fable 5's numbers come from Anthropic. GPT-5.6 Sol's come from OpenAI. The one row that is worth more than the rest combined is GDPval-AA v2 — because Artificial Analysis ran it independently across all three models.

Results

Agent GDPval-AA v2 Terminal-Bench 3.0 SWE-bench Pro Notes
GLM-5.3 1769 28.3 (vendor) ~67 (est.) Same base as GLM-5.2; all gains from post-training. Weights not yet open.
Claude Fable 5 1743 84.6 (vendor) 80.0 (vendor) Strongest on deep engineering work. Fallback caveat noted by Alibaba.
GPT-5.6 Sol 1730 88.8 (vendor, max) 64.6 (vendor) Leads terminal tasks but trails on SWE-bench Pro.

Key findings

  1. GLM-5.3 wins the independent row. GDPval-AA v2 at 1769 puts it ahead of both Western flagships — and this is the one number not run by the model's own vendor.
  2. Terminal-Bench 3.0 jump is real but vendor-reported. GLM-5.3 went from 4.6 to 28.3 on Terminal-Bench 3.0 — a massive jump, but Z.ai ran this benchmark themselves. Independent replication pending.
  3. Fable 5 still leads on deep engineering. SWE-bench Pro at 80.0 is the honest weak spot for both GLM-5.3 and GPT-5.6 Sol. If your agents do deep codebase work, Fable 5 remains the pick.
  4. GPT-5.6 Sol leads on terminal tasks. Terminal-Bench 2.1 at 88.8 (max) is the highest score, but this is a vendor-reported max configuration.
  5. Price matters. GLM-5.3 at $0.80/$3.00 per 1M tokens is roughly 6x cheaper than GPT-5.6 Sol and 10x cheaper than Fable 5. The independent benchmark lead at that price point is remarkable.

Caveats

  • Weights not yet open: GLM-5.3's weights are expected end of August 2026. Until then, API-only — and API results may differ from self-hosted.
  • Vendor benchmarks dominate: Only GDPval-AA v2 is independently scored. Terminal-Bench and SWE-bench numbers are vendor-run and may involve different harnesses, tool configurations, and sampling.
  • Fallback caveat: Alibaba's own benchmark table notes Fable 5 results "may involve fallbacks" — this affects the validity of the SWE-bench Pro comparison.
  • Pass-k not measured: These are single-run scores. Pass^4 reliability — the metric that matters for production — is not yet available for GLM-5.3.

When to use this comparison

Use this benchmark to decide which coding agent to try for your workflow. For terminal-heavy agent tasks at low cost, GLM-5.3 is the leading candidate. For deep codebase engineering, Fable 5 remains the pick. For maximum terminal performance regardless of cost, GPT-5.6 Sol leads.

Do not use this as your only evaluation. Run your own tasks through each model before committing — vendor benchmarks and even independent benchmarks do not predict your specific workload perfectly.