AI Token Plans Explained: How Chinese Labs Are Reinventing the Cost of Coding Agents

A new pricing model is quietly reshaping how developers pay for AI-assisted coding. As agentic workflows—where an AI runs long, tool-heavy loops largely on its own—become the default way engineers ship code, a cluster of Chinese AI labs has introduced flat-rate "token plans" that sit between two uncomfortable extremes: threadbare free tiers and the unbounded bills of pay-per-million-token APIs.

MiniMax, Xiaomi's MiMo team and Zhipu's GLM division are all now selling monthly subscriptions that grant a rolling usage quota consumed inside coding tools such as Claude Code, Cline and OpenCode. The pitch is simple: predictable cost. For developers who have watched an autonomous agent quietly burn through hundreds of dollars in an afternoon, that predictability is the entire point.

What a token plan actually is

A token plan is a flat monthly fee that buys a rolling quota of model usage—not an unlimited buffet, and not a metered tab. Three traits define the category. First, rolling quota windows: most plans meter usage across a 5-hour rolling window plus a weekly cap (MiniMax, GLM), or a monthly credit pool (Xiaomi MiMo). Second, a hard stop rather than silent overage—when quota is exhausted, service halts by default instead of billing you into oblivion. Third, a tool-restricted subscription key that works only inside approved coding agents, not for arbitrary API or production traffic; providers explicitly reserve the right to ban keys used for automated scripts or backends.

The contrast with traditional API pricing is stark:

These plans are proliferating now because agentic loops consume tokens at a rate that makes per-token billing genuinely unpredictable, and buyers are hunting for a fixed ceiling. Search interest in "token plan" has climbed accordingly.

The three plans at a glance

[IMAGE_1]

Figures are dated snapshots; promos and version numbers move fast.

MiniMax: the multimodal outlier

MiniMax is the only plan of the three that is explicitly full-spectrum multimodal—one subscription covers language, speech, video, music and image through a shared usage bar. It evolved from the lab's former "Coding Plan." Tiers run Plus ($20) for prototyping, Max ($50) for daily coding, and Ultra ($120) for heavy agent workflows, all granting every model and differing mainly by quota size and concurrent agent count. A recent upgrade shifted the meter from a fixed request count to actual token consumption. At the limit, users can cover with purchased credits, upgrade instantly, swap in a pay-as-you-go API key to continue uncapped, or wait for reset.

MiniMax also carries the category's most-cited real-world hazard. In a widely upvoted community warning, developers reported that prompt-caching on MiniMax's Anthropic-compatible endpoint is broken: standard cache instructions from IDE extensions are ignored, so the backend re-reads the full project history every turn at full price. One user reported burning roughly 300 million tokens in 36 hours of light coding.

The community-sourced fix: switch the endpoint from Anthropic to OpenAI-compatible (https://api.minimax.io/v1), after which caching reportedly works and quota lasts far longer.

This remains user-reported and is not confirmed by MiniMax, but it is corroborated across multiple accounts—alongside complaints of a laggy first-party Mac app and unresponsive support.

Xiaomi MiMo: the cheapest door in

MiMo offers the lowest entry at $6/month and uses a monthly credit pool rather than rolling windows. Tiers scale from Lite (roughly 4.1 billion credits) up to Max ($100), with annual plans providing roughly twelve times the monthly volume. Its lineup spans about six models—including mimo-v2.5-pro, the base mimo-v2.5, an ASR model and several TTS voices—with TTS reportedly free for now and ASR billed by audio duration.

The pro model consumes roughly three times the credits of the base model, making model choice a serious cost lever, and off-peak overnight Beijing hours attract a consumption discount. MiMo will not drain any bonus or balance at the hard stop—but there are no refunds on unused credits, and its coding-tools-only rule is enforced explicitly: using the key for scripts or custom backends is a violation that can get it banned.

GLM Coding Plan: cheap high quota, with an asterisk

Zhipu's GLM plan is the cheapest high-quota agentic option tracked, and its coding model is a strong long-context open-source coder. It is coding-only. Tiers run Lite ($18), Pro ($72) and Max ($160), scaling quota roughly 1×, 5× and 20×, with promotional pricing around 30% off. Quota is denominated in prompts—one user turn each—rather than tokens.

The catch is a multiplier system: flagship models consume roughly 3× during peak hours and 2× off-peak, softened by a limited-time 1× off-peak promotion. The plan is a pure hard stop with no overage. Community sentiment is mixed: one Lite user reported a single 14-minute multi-agent review devouring half a 5-hour window, and several argue raw API remains cheaper for comparable models until GLM's quotas roughly double.

OpenCode Go: the dollar-cap wildcard

OpenCode Go takes a different tack. At $10/month ($5 first month), its limits are dollar-denominated—$12 per 5 hours, $30 weekly, $60 monthly—not counted in prompts or requests. Its headline feature is model freedom: one key unlocks 13–14 open models across GLM, Kimi, MiMo, MiniMax, Qwen and DeepSeek families, with no single-vendor lock-in. Because the cap is a dollar figure, a cheap model can yield tens of thousands of requests per window while a premium coder yields only hundreds. Hosted internationally with a zero-retention policy, it appeals to proprietary work. A popular pattern pairs a direct vendor subscription as the main agent with Go for sub-agents and model variety.

Which plan fits which developer

[IMAGE_2]

Before you buy

Quota units are not comparable at face value—tokens, prompts, credits and dollars measure different things. Watch reset cadence: rolling windows can strand you mid-task, while monthly pools cannot. Confirm whether a plan halts or bills at the limit; only MiniMax continues silently once a pay-as-you-go key is enabled. Respect the tool restrictions, and schedule heavy runs off-peak where multipliers apply.

For light or bursty work, a raw API is often still cheaper—subscriptions only pay off once daily usage climbs past roughly a thirtieth of the monthly cap. In practice, a $50 plan starts to win once you would otherwise spend about $1.70 a day on API calls; below that, metered billing is the cheaper bet.

The token plan is not a universal upgrade. But for the developer whose agent runs all day, it converts a source of billing dread into a line item they can actually forecast—and that alone explains why the category is spreading fast.