'Base usage allowance' is the industry's new 'unlimited*'

What happened

ZCode, Z.ai's desktop harness for GLM-5.2, hit 410 points on Hacker News this week. Most of the discussion focused on the obvious: it looks like a Codex clone (paxys: "UI-wise this looks a lot closer to Codex than Claude Code. It's basically an exact copy"), it isn't open source (seizethecheese: "somewhat surprised that this is not open source"), and it interoperates with the usual CLI agents via documented adapters.

Buried in the thread, though, was the more interesting observation. Commenter cube00 flagged the pricing page's language directly: "base usage allowance included" on the entry tier, with higher plans sold as multipliers of that base — but nowhere does Z.ai define what the base is. No token count, no request count, no compute-hour figure. Just a marketing noun standing in for a spec.

ZCode is not the outlier. It is following the template.

Why it matters

Walk the pricing pages of the tools a senior developer actually uses in 2026 and the pattern is impossible to miss. Cursor's Pro tier now sells "included usage" with Ultra branded as "20x usage" — 20x of what is never specified in a number, only referenced as "our standard limits." GitHub Copilot Business and Enterprise are tiered on "premium requests," a unit Microsoft coined in 2024 that maps to model calls but whose exact ratio to a chat turn, an agent step, or an inline completion is documented only in a support article that has been quietly revised at least three times. Claude Code Pro and Max are marketed as "5x" and "20x" of the base — Anthropic's own documentation states the base "varies based on demand" and points at a rate-limit page that no longer publishes numeric ceilings. And now ZCode: "base usage allowance included," "expanded allowance," "maximum allowance." Four of the biggest AI coding tools on the market, one shared vocabulary, zero disclosed baselines.

The reason this vocabulary spread so fast is that it solves a real vendor problem. Inference costs are volatile — GPU spot pricing swings 30-40% week to week, model providers renegotiate wholesale rates, and long-context workloads are drastically more expensive than short ones. If your pricing page publishes a hard token cap, every margin squeeze becomes a public pricing change, which triggers churn, press coverage, and support-ticket floods. If your pricing page publishes an allowance, you can quietly retighten the definition of that allowance and your only tell is a Discord thread of confused users comparing notes.

We already have receipts on this dynamic. In April 2026, Cursor users hit a wall where Ultra plans that had comfortably handled agent runs the prior month began throwing rate limits at similar workloads; Cursor's response was that "limits are dynamic and adjust to fair use," which is exactly what the vague terminology enables them to say. GitHub Copilot's premium-request definition was tightened in Q1 2026 to count agent tool calls that had previously been bundled, a change communicated via a blog post rather than a plan-change notification. Anthropic has been the most transparent of the group, publishing weekly rate-limit adjustments — but even there, the numeric ceilings visible in 2024 are gone from public docs.

This isn't a conspiracy; it's a market equilibrium. When one vendor moves to opaque units, the vendors who publish hard numbers become the ones eating the volatility, so they either eat their margins or match the language. ZCode matched the language on day one.

What this means for your stack

If you are the person who signs the AI coding tool invoice for your team, three things follow.

First, treat every "allowance" tier as variable-rate. The plan you priced into your Q3 budget can materially degrade without your vendor changing a single line of published pricing. Build a monthly review into your finance process the same way you would for a cloud spend that's subject to on-demand instance pricing. If your team's velocity depends on the tool, the exposure is real.

Second, instrument your own usage. You cannot negotiate on an allowance you cannot measure. Every major agent CLI (Claude Code, Codex, Aider, and yes ZCode's documented adapters) emits token counts per turn. Pipe them into whatever observability stack you already have. When your vendor tightens the definition of "base usage," you want to see the degradation as a graph, not as a Slack complaint from your staff engineer.

Third, keep an open-weights escape hatch warm. The whole reason vendors can get away with allowance opacity is that switching costs are high — muscle memory in a specific CLI, prompts tuned to a specific model, workflow integrations. If you're heavily invested in one closed harness, do a quarterly dry run against an open equivalent (OpenCode, Aider against a locally-hosted GLM or a smaller open model). The dry run itself does not need to be a migration. Its purpose is to keep the option credible so that when a vendor's allowance definition drifts, you have a real BATNA rather than a theoretical one.

Looking ahead

The interesting question is how long this equilibrium holds. Enterprise buyers with real procurement teams will eventually demand numeric SLAs — Fortune 500 legal departments do not sign contracts that say "base allowance may vary based on demand." Expect the first crack to be an enterprise-tier pricing page that quietly publishes actual token ceilings while the consumer-tier page continues to sell "allowances." ZCode's arrival with the same vocabulary as everyone else is a signal that this is now the default, and the default is what enterprise buyers eventually push against. Watch for the vendor that breaks ranks first — they'll be the one that just closed a nine-figure enterprise contract and needed a number to put on paper.