CommandCode Model Efficiency Report

• Date: 2026-09-08
• Scope: Performance-per-cost analysis of the CommandCode Provider API model catalog, prompted by GLM-5.3-Flash being the daily-driver model in Hermes (and @smol role in omp).
• Method: SearXNG web search (self-hosted instance) → Artificial Analysis model pages (Intelligence Index v4.3, list prices, throughput) scraped via curl and parsed programmatically, supplemented by OpenRouter pricing pages, the Ampere.sh GLM-5.3-vs-Flash comparison, and vendor docs (Z.ai, Hugging Face).

Verdict

Keep GLM-5.3-Flash as the daily driver. It is already the most cost-efficient model on the CommandCode catalog — by a wide margin, not a marginal one. No switch recommended.

Efficiency Table

Sorted by Artificial Analysis' full Intelligence Index evaluation suite cost per Index point (their apples-to-apples efficiency measure). All prices are list prices per 1M tokens as tracked by Artificial Analysis (first-party serving, Sep 2026).

[^1]: II = Artificial Analysis Intelligence Index v3.4.3-class score as published on each model page (higher is better; page-stated medians ranged 17–24 at fetch time). [^2]: Blend = 3:1 input:output token mix, a standard workload approximation.

GLM-5.3-Flash is roughly 8x cheaper per quality point than its own flagship (GLM-5.3) and ~20x cheaper per point than GPT-5.5.

Findings

1. GLM-5.3-Flash profile

• Intelligence Index 42 (well above the class median of ~17); 1M context window; 60 tok/s output.
• Natively multimodal (vision) — the only vision-capable GLM-5.3 variant; the standard GLM-5.3 is primarily text-based. Vendor docs list input modality as Video / Image / Text / File.
• Sparse architecture: 320B total parameters with only 18B active per token, which is why it is this cheap to serve.

2. Coding gap to the flagship is small; the price gap is ~9x

Ampere.sh (Aug 2026) side-by-side, both at reasoning effort:

• Flash is slightly ahead on agentic automation (AutomationBench), and the Terminal Bench / DeepSWE deficits are a few points, not a collapse.
• Ampere's own conclusion: "whether those extra few benchmark points justify paying roughly nine times more per token… For many production coding agents, they may not."
• Flash additionally wins for visual coding / browser-use / computer-use workflows because it can see.

3. If you ever want more headroom

• GLM-5.3 (II 45, +3 pts) is the only sensible upgrade on this catalog. Everything else is worse value per point.
• Kimi K3 (II 44) at $3/$15 is the worst value per point after GPT-5.5 — despite similar raw intelligence.
• Qwen3.8 Max (II 40) is ~2x flash's blended price for 2 points less intelligence, at roughly half the speed (40 vs 60 tok/s).
• GLM-5.2 is strictly dominated by GLM-5.3-Flash (same price as flagship, 3 fewer II points, slower).

4. Speed alternative

DeepSeek V4 Flash is the throughput king of the catalog (126 tok/s, rank 7/112 on AA's speed chart) at 2.8x flash's cost per point but still cheap in absolute terms. Consider it only for long high-volume batch jobs where wall-clock time dominates token cost.

Caveats

• Prices are AA-tracked first-party list prices. On CommandCode you pay plan quota, not per-token — what actually matters for quota burn is tokens consumed per task. GLM-5.3-Flash is also competitive here: 180M tokens generated on AA's suite (verbose, rank 25/112) versus 210M for GLM-5.3 and 240M for DeepSeek V4 Flash.
• Verbosity at max effort is flash's one inefficiency. If quota burn ever becomes a concern, dropping trivial calls from max to high effort costs little quality. (Note: the pi-commandcode-provider catalog for z-ai/glm-5.3-flash exposes effort levels low, high, max only — there is no xhigh.)
• CommandCode model availability rotates. This reflects the catalog as of 2026-09-08; re-run the scrape before relying on it for a purchase decision.
• Intelligence Index scores between differently-benchmarked pages (different class medians) are approximately comparable only.

Current configuration (verified working, no change needed)

References

• Artificial Analysis — GLM-5.3-Flash (Intelligence, Performance & Price): https://artificialanalysis.ai/models/glm-5-3-flash
• Artificial Analysis — GLM-5.3: https://artificialanalysis.ai/models/glm-5-3
• Artificial Analysis — DeepSeek V4 Flash 0731 (Reasoning, Max Effort): https://artificialanalysis.ai/models/deepseek-v4-flash
• Artificial Analysis — DeepSeek V4 Pro 0813 (Reasoning, Max Effort): https://artificialanalysis.ai/models/deepseek-v4-pro
• Artificial Analysis — Qwen3.8 Max: https://artificialanalysis.ai/models/qwen3-8-max
• Artificial Analysis — Kimi K3 (max): https://artificialanalysis.ai/models/kimi-k3
• Artificial Analysis — GPT-5.5: https://artificialanalysis.ai/models/gpt-5-5
• Artificial Analysis — Gemini 3.5 Flash: https://artificialanalysis.ai/models/gemini-3-5-flash
• Artificial Analysis — GLM-5.2: https://artificialanalysis.ai/models/glm-5-2
• Ampere.sh — GLM 5.3 vs GLM 5.3 Flash: Benchmarks, Price & Coding (Aug 27, 2026): https://www.ampere.sh/blog/glm-5-3-vs-glm-5-3-flash
• Z.ai Developer Docs — GLM-5.3-Flash overview: https://docs.z.ai/guides/llm/glm-5.3-flash
• Hugging Face — zai-org/GLM-5.3-Flash model card: https://huggingface.co/zai-org/GLM-5.3-Flash
• OpenRouter — GLM-5.3 pricing cross-check: https://openrouter.ai/z-ai/glm-5.3
• OpenRouter — Kimi K3 pricing cross-check: https://openrouter.ai/moonshotai/kimi-k3
• Ampere.sh — GLM 5.3 Flash vs GPT-5.4 Mini: https://www.ampere.sh/blog/glm-5-3-flash-vs-gpt-5-4-mini

Scraped metrics and computed efficiency ratios (blend pricing, suite-cost-per-II-point) are in the session scratch files under /tmp/aa-.json; the raw Artificial Analysis page text is preserved in /tmp/aa-raw.json for audit.*