Model Comparison and Evaluation in HagiCode

Scope and method

• Goal: Provide model-selection guidance based on real integration experience in HagiCode.
• Task types: Frontend component implementation, backend API refactoring, test completion, and documentation generation.
• Evaluation axes: delivery effectiveness (can it reliably finish real tasks) and cost-effectiveness (cost + domestic availability).

Test-time and scenario notes

• Latest test date: 2026-03-08
• Test period: 2026-03-01 to 2026-03-08
• Sample basis: Subjective evaluation from real HagiCode engineering workflows, not vendor benchmark numbers.
• Applicability: Conclusions are scoped to this project’s current workflow and constraints.

This page only lists tested models

The following models were actually integrated and used by our team:

• GLM 4.7
• GLM 5
• Qwen 3.5
• Qwen Code Next
• GPT 5.3 Codex
• GPT 5.4
• Minimax M2.5

Comparison snapshot (tested models)

Model	Test Date	Delivery effectiveness	Cost-effectiveness	Primary experience
GPT 5.4	2026-03-08	Very high	Medium-high	Frequently exceeds baseline requirements with strong engineering quality
GPT 5.3 Codex	2026-03-08	Very high	Medium-high	High completion quality within scope, strong engineering output
GLM 5	2026-03-08	High	High	Stable overall performance for our requirements
GLM 4.7	2026-03-08	High	Very high	Reliable delivery with better cost control
Minimax M2.5	2026-03-08	Medium-high	Highest	Can achieve most goals, but code-closing errors happen more often
Qwen 3.5 / Code Next	2026-03-08	Medium	Medium-high	Lower completion ranking in our scenarios

Delivery-effectiveness ranking (author recommendation)

Ranked by task completion quality and engineering practice quality:

1. GPT 5.4
2. GPT 5.3 Codex
3. GLM 5
4. GLM 4.7
5. Minimax M2.5
6. Qwen (3.5 / Code Next)

Key findings

• Except Qwen, all other tested models can achieve our target outcomes to some degree.
• GLM 4.7+ (GLM 4.7 and GLM 5) generally completes our requirements smoothly.
• GPT 5.3 Codex and GPT 5.4 not only complete requirements but also produce better engineering practices and implementation quality.
• Minimax M2.5 has a recurring weakness: code-closing errors (e.g., incomplete bracket/block closure), so extra review is needed.

Cost-effectiveness ranking (cost + domestic availability)

Ranked by cost-to-benefit ratio and whether the model can be used smoothly in China:

1. Minimax M2.5
2. GLM 4.7
3. GLM 5
4. Qwen 3.5 / Code Next
5. GPT 5.3 Codex
6. GPT 5.4

Note: this ranking differs from the delivery-effectiveness ranking; it mainly reflects the practical constraints of economic cost and availability.

Selection guidance

• Quality-first: choose GPT 5.4 / GPT 5.3 Codex.
• Balanced strategy: choose GLM 5 / GLM 4.7.
• Cost-first: choose Minimax M2.5 (with stricter code-closure checks).
• Practical routing: use premium models for critical tasks and cost-efficient models for routine tasks.

Untested-model statement

For models not listed here, we currently have no test data and no hands-on experience, so we do not provide evaluations.

If sponsors provide access to additional models, we will run experience-based evaluations in our real workflow and update this page.