Zhipu AI, one of China’s largest AI labs, released ZCode on July 6, a coding agent environment built on its open-source GLM-5.2 model that positions itself as a direct, cheaper alternative to Anthropic’s Claude Code and OpenAI’s Codex. The tool handles code writing, debugging, testing, code review, and Git workflows through natural language commands, according to The Decoder.

ZCode’s dedicated agent operates within a 1M-token context window, allowing multi-step programming tasks without losing thread. The agent can also be controlled remotely through Feishu, WeChat, or mobile apps, embedding coding agent capabilities into the enterprise communication platforms where Chinese development teams already work.

The Pricing Gap

The competitive angle is blunt: cost. GLM-5.2 runs at $1.40 per million input tokens and $4.40 per million output tokens, according to Zhipu’s official pricing. Claude Opus 4.7, the model powering Claude Code’s most capable mode, costs $5 input and $25 output. GPT-5.5 costs $5 input and $30 output. That puts GLM-5.2 at roughly 72% cheaper on input and 82% cheaper on output than Opus.

New ZCode customers get a free five-day trial with up to 5 million tokens per day, according to ZCode’s documentation. Paid subscribers receive approximately 50% additional token quota through July 2026.

Performance: Close, Not Equal

The price gap only matters if the model performs. A hands-on comparison by Snowflake CEO Sridhar Ramaswamy across 103 coding tasks found GLM-5.2 and Opus 4.7 “nearly tied” after three attempts per task, with completion rates of 66% versus 67%, according to The Decoder’s analysis of Ramaswamy’s findings.

The gap widens on first-attempt accuracy: Opus hit 53.7%, GLM only 47.6%. GLM also consumed nearly double the tokens per task (860 million versus 439 million across the benchmark), which eats into the pricing advantage. On one task, GLM fired off 411 tool calls in 24 minutes and still failed all three attempts, while Opus solved it with 49 calls in 9 minutes.

GLM-5.2 shipped in June 2026 under an MIT license. On the FrontierSWE benchmark for long-horizon engineering tasks, it scored 74.4%, one point behind Opus 4.8 and slightly ahead of GPT-5.5, according to The Decoder. On ultra-long-horizon benchmarks like SWE-Marathon, the gap widens significantly, with GLM-5.2 reaching only half of Opus 4.8’s score.

The Geographic Bet

ZCode represents a specific strategic play: a Chinese vendor marketing aggressively to Western developers. The question for teams evaluating it is whether the 70-80% cost savings justify the tradeoffs in first-attempt reliability and token efficiency, particularly when data residency and security review requirements add friction for US and EU organizations routing code through Chinese-hosted infrastructure.