GLM-4.7 Beats GPT-5 in Coding Tests at One‑Seventh the Cost

At the end of 2025, Zhipu released GLM-4.7, a 358‑billion‑parameter open‑source model marketed as "your new coding partner".

Core Strength: Coding Performance

On the SWE‑bench software‑engineering benchmark, GLM-4.7 achieved a score of 73.8% , a gain of nearly 6 percentage points over GLM-4.6 and up to 12.9% on multilingual programming tasks. On Terminal Bench 2.0, its performance rose by 16.5% , indicating more accurate command‑line operations.

In practical terms, the model can handle complex software‑engineering problems beyond simple code generation.

Vibe Coding: Better UI/UX Output

GLM-4.7 adds a feature called Vibe Coding that optimises generated UI code for modern, clean designs and precise layout sizing. The model also produces well‑formatted PPT slides. Example outputs include dark‑mode websites and artistic particle‑effect pages, which the author describes as "god‑level" for front‑end developers.

Thinking Enhancements

The model introduces Interleaved Thinking , which performs a deep logical reasoning step before each operation or tool call, avoiding blind trial‑and‑error. It also supports Preserved Thinking , remembering previous reasoning across multi‑turn dialogues to maintain stability on long tasks.

Benchmark Comparison with Top Models

AIME 25 (Math Competition) : GLM-4.7 scores 95.7 , surpassing GPT‑5 (94.6) and Claude Sonnet 4.5 (87.0).

GPQA‑Diamond (Expert Q&A) : GLM-4.7 scores 85.7 , higher than Claude Sonnet 4.5 (83.4).

HLE (Human Last Exam) : GLM-4.7 achieves 42.8% , far above Claude Sonnet 4.5's 32.0%.

Across most core leaderboards, GLM-4.7 ranks in the first tier of current AI models and leads in several domains.

Open‑Source, Affordable, Ready to Use

GLM-4.7's weights (358B parameters) are published on HuggingFace, allowing local deployment and full data control.

API pricing is roughly 1/7 of Claude Sonnet 4.5, offering a better cost‑performance ratio.

Open‑source : Model weights available for download.

Cheap : API cost is one‑seventh of the main competitor.

Compatibility : Supports major agent frameworks such as Claude Code and Roo Code; existing GLM Coding Plan subscribers can upgrade immediately.