GLM-5.2 from Chinese startup Z.ai beats GPT-5.5 on coding at 1/6th cost

Z.ai (formerly Zhipu AI), a Beijing-based startup, released GLM-5.2, a 753-billion-parameter open-weight model that posts stronger performance than GPT-5.5 on several long-horizon coding benchmarks while costing roughly one-sixth as much. On SWE-bench Pro, GLM-5.2 scored 62.1 versus GPT-5.5's 58.6; on FrontierSWE (simulating hours-long engineering tasks), it hit 74.4% versus GPT-5.5's 72.6%. On Design Arena's crowdsourced leaderboard, GLM-5.2 ranked #1 with an Elo of 1360, beating Claude Fable 5. The model costs $1.40 input / $4.40 output per million tokens versus GPT-5.5's $5 input / $30 output.

GLM-5.2 extends a stable 1-million-token context and introduces IndexShare, a sparse-attention technique that reuses indices across transformer layers to cut per-token compute by 2.9x at long context windows. It is released under an unrestricted MIT license, allowing enterprises to download weights, fine-tune, and self-host without API dependencies. Independent benchmarks (Semgrep) found GLM-5.2 beat Claude Code on IDOR vulnerability detection at $0.17 per bug found.

The release marks the first open-weight model to genuinely narrow the frontier gap on production coding tasks, directly pressuring closed-model economics. Snowflake's CEO tested GLM-5.2 against Opus 4.7 on internal benchmarks and found it competitive despite higher token consumption. With OpenAI and Anthropic models restricted by government review and priced at premium rates ($30–$50/M tokens), Z.ai's aggressive pricing and MIT licensing create a cost-quality frontier that enterprises deploying custom models at scale will find hard to ignore.