Z.ai pays for ZCode users in tokens, not cash: 100 million each to 50,000 signups

The advertised ceiling is just the two offer numbers multiplied: 100 million tokens across 50,000 seats comes to 5 trillion ^[1]. It binds only if every allocation is claimed and then spent, and Z.ai has given itself two days to fill the cohort ^[5]. Set against the 1 million ZCode users the company reported on August 11 ^[9], the 50,000 seats amount to about 5 percent of that self-reported base ^[2]. This is a sampling exercise sized to a deadline, not a growth push.

The number that matters to a working engineer is 100 million, not 5 trillion. Coding agents burn inference because they read the repository again, run the command again and revise their own output ^[16], and Z.ai's own framing is that the allowance is meant to carry a longer project rather than a handful of prompts, inside an interface where the company controls onboarding, model selection and the route to a paid plan ^[17]. For someone already holding a Cursor, Claude Code or GitHub Copilot seat ^[19], that turns an evaluation from a budget conversation into a signup.

The benchmark case for bothering is thinner than the headline figure suggests. Z.ai claims a 50 percent gain on its internal coding benchmark ^[12] and reports Terminal-Bench 3.0 rising from 4.6 to 28.3 ^[13] and DeepSWE v1.1 from 46.2 to 66.9 ^[14]. The DeepSWE move is 20.7 points, a 44.8 percent relative gain, which is the one that actually tracks the 50 percent claim ^[4]. Terminal-Bench 3.0 is a 6.2x multiple ^[3], flattering mostly because the starting point was near the floor, and 28.3 still leaves 71.7 points on the table ^[5]. All of it is Z.ai's own evaluation and has not been reproduced independently ^[15]. GLM-5.3 also sits on the same base as GLM-5.2 with additional post-training ^[11], which fits a company with serving capacity to spend rather than a fresh pretraining bill to amortise.

The seams show in the product surface. Z.ai announced GLM-5.3 on August 14, 2026 without a documented access path, then opened it through the API ^[10], while the ZCode product page still promotes deep GLM-5.2 integration even as this promotion pairs the agent with GLM-5.3 ^[7]. Anyone spending a free allowance should confirm which model is actually answering the session, because the marketing and the product page currently name different ones ^[7]. Remote job control through WeChat, Feishu or Telegram ^[8] is a real feature and a reminder whose workflow the agent was shaped around.

Then there is the load. Z.ai's own reading is that heavy use across all 50,000 allocations puts its serving layer on trial ^[18]. That risk is asymmetric in a way cash discounting is not: a queue or a timeout arrives during the first sustained session, which is the session the whole campaign is designed to produce. And because the outlay is denominated in model usage rather than dollars ^[3], there is no figure to compare against a competitor's discount or against Z.ai's own margin. The company that grew out of Tsinghua research by publishing models, building a developer base and selling the infrastructure underneath ^[22] is now testing whether that infrastructure holds when the models are free.