Zhipu just turned Anthropic's worst week into a recruiting pitch, dangling free tokens and fatter data quotas right as developers were still absorbing news that Claude Code had been quietly flagging Chinese users.
On Wednesday, Beijing based Zhipu AI, listed in Hong Kong as Knowledge Atlas Technology and known abroad as Z.ai, released ZCode, a harness that lets its GLM-5.2 model run as an autonomous coding agent. The timing wasn't subtle. According to a report from the South China Morning Post, the launch landed just as Anthropic was pulling hidden code that had been covertly tracking whether Claude Code users were tied to Chinese AI labs.
To pull developers away from Cursor, GitHub Copilot and Anthropic's own Claude Code, Zhipu sweetened the deal. Existing subscribers got a 50 percent bump in their data quotas. New ZCode users got 5 million free tokens, the basic units of data an AI model consumes, plus a five day trial that needs no credit card. A separate promotion running through July 31 meters token use at 0.67x during off peak hours and 2x at peak, which works out to roughly 1.5 times the effective quota for anyone willing to code at odd hours.
The offer sits on top of a model that was already turning heads. GLM-5.2, a 753 billion parameter system Zhipu open sourced under an MIT license last month, scores 62.1 on SWE-bench Pro against GPT-5.5's 58.6, according to benchmarks reported by VentureBeat. On FrontierSWE it hits 74.4, edging past GPT-5.5's 72.6 and landing within a point of Claude Opus 4.8's 75.1. Z.ai prices API access at $1.40 per million input tokens and $4.40 per million output tokens, a fraction of what GPT-5.5 costs to run, with subscription plans starting around $18 a month.
The model's rise had already moved markets. Zhipu's Hong Kong listed shares jumped after GLM-5.2's release, and Tom's Hardware reported the open weight model climbing to the top of open weight rankings while running on Huawei silicon, a detail that matters given the export controls still limiting Chinese access to Nvidia's best chips. Some in Silicon Valley have already called it another DeepSeek moment, the same phrase used last year when a different Chinese lab upended assumptions about how much frontier AI has to cost.
The backdrop here matters more than usual. A Reddit user going by LegitMichel777 reverse engineered Claude Code on June 30 and found code that had sat unnoticed since version 2.1.91, released back on April 2, with no mention in any release notes. If a user's timezone read as Chinese, the software swapped the date format from dashes to slashes and replaced an ordinary apostrophe with one of three visually identical Unicode characters, depending on which flags fired. None of it was visible to a human reading the screen. Portions of the detection logic were XOR obfuscated with the key 91, according to Semafor's reporting on the rollback.
Anthropic engineer Thariq Shihipar said on X that the code was
Also read: Amazon is quietly building the AI chips that power your Echo • Macron and Modi Are Personally Courting Tech CEOs to Win the AI Infrastructure Race • Startups Are Racing to Put AI Data Centers in Orbit Before Big Tech Gets There