GLM-5.2 vs. Kimi K2.7: Why GLM Wins the Code Reliability Test

When two of the strongest open models hit the market within days of each other last June, one question dominated the conversation: which one actually delivers for real-world coding tasks? According to five independent evaluations, Z.ai's GLM-5.2 edges out the competition, though rarely with ease.

In quick tasks, the two models traded wins. In tasks that required a second look, a difference emerged.

In quick tasks, almost a tie

Fahd Mirza ran both within the Hermes agent against a goal-difference tie-breaker bug in a World Cup application. Both models diagnosed the error and built a new 32-team bracket in a single prompt. Kimi finished faster — five minutes — and added a tournament progression detail on its own. Mirza scored the two as almost tied.

Samuel Gregory reached the same conclusion upon seeing Kimi's agent swarm propagate a redesign across an entire site. He called both fantastic and placed them close to Opus 4.5 in reliability. A third evaluator split a sorting visualizer task: he preferred Kimi's design and GLM's functionality — before watching both build a functional Rust file application from a single prompt, although Kimi delivered an error in the dependency file that required manual correction.

On the second look, GLM pulled away

Web3 Wesley ran both on three tasks — a site, a game, and social media text — and handed the result to Claude to inspect. Kimi's coffee subscription site looked more polished at first glance. The code review found something different: a broken mobile menu and a game whose difficulty level came from a missing delta-time calculation, not intentional design. Wesley gave GLM the win in two of the three tests, with a tie on the text.

"Kimi's site looked prettier — until the code was inspected. The mobile menu was broken and the game's difficulty was accidental."

— Web3 Wesley, independent evaluator

The Better Stack evaluation reached a similar split. GLM produced a Three.js racing game in a single prompt, using about 40,000 tokens. Kimi needed a follow-up prompt and consumed about 110,000. In a full financial dashboard, GLM connected a Next.js and Prisma stack without errors. Kimi set up a React and Express structure writing to a local SQL file — a choice the evaluator judged less scalable.

What each model does best

What the numbers say about the models

GLM-5.2 carries 744 billion parameters with 40 billion active, MIT license, and a one-million token context window. Artificial Analysis named it the highest-scoring open model on its intelligence index this month — 51 points, 11 more than GLM-5.1. It outperformed GPT-5.5 on the GDPval benchmark and took first place in the Design Arena in single-turn HTML web design, the first model to cross the Claude line in that category, including Fable 5.

Kimi K2.7 Code is the more specialized instrument of the two: one trillion parameters with 32 billion active, always-on reasoning mode, and the only one capable of reading images. Its 256,000-token context window was flagged by several evaluators as limiting for production code.

In cost, GLM also pulled ahead. Artificial Analysis recorded an average of US$ 0.50 per task — the lowest cost at this intelligence level. The Better Stack evaluator said he could replace Sonnet or Opus in simpler jobs without noticing a difference. Moonshot lists Kimi K2.7 Code at US$ 0.75 per million input tokens and US$ 3.50 per million output, with subscription plans starting at US$ 15 monthly.

For developers choosing between the two Chinese open models: evaluators gave GLM-5.2 the slight edge for code in general. Kimi K2.7 Code is the choice when the task needs image input or the agent swarm.

No result here comes from a standardized benchmark. They are five evaluators, with their own prompts, evaluating in their own way. What matters is not any isolated test — it is the pattern that appears in all five.