Chinese AI Coding Plan Pricing in 2026: All 7 Providers, Domestic vs Overseas, Benchmarks, and Honest Buying Advice

Token is the new data plan

Remember how mobile data plans evolved? In 2009, you paid 100 yuan for 1 GB and thought it would last forever. By 2013, the same money barely bought a few gigabytes. Then came “unlimited” plans — except they throttled you after a threshold. By 2020, truly unlimited data was gone, and everyone was counting gigabytes again.

Token pricing is following the exact same trajectory. From late 2024 through mid-2025, providers slashed per-million-token prices to fractions of a yuan — cheap enough that you did not bother checking usage. From late 2025 onward, the wind shifted: purchase limits, price hikes, discontinued low tiers, and first-purchase discounts that turned from “buy anytime” to “10:30 AM daily flash sale.”

You thought tokens were air. They are actually data plans — the free stuff is disappearing, and the cheap stuff is getting more expensive.

I am an engineer who consumes over a billion tokens per month. From Claude Code to Cursor, from agent swarms to automated pipelines, tokens are oxygen — you do not notice them until they run out. What follows is my honest, page-by-page audit of every major Chinese AI coding plan and its overseas counterpart, written for fellow engineers who need to make smart buying decisions right now.

The 30-second summary

Look at that table carefully. Some providers charge more overseas but give you more quota. Others are simply more expensive internationally with no upside. The right answer depends on your payment method, your network latency tolerance, and how much quota you actually burn per session.

Zhipu GLM: two rounds of price hikes, and domestic users still get a bargain

At an 8:1 exchange rate, the overseas Max tier costs roughly ¥1,280/month while the domestic Max is only ¥469. Same model, same quotas — mainland users pay less than half what overseas users pay.

But the overseas version has one crucial advantage: no purchase limits. The domestic Max tier frequently sells out, with only about 20% of quota released daily at 10:00 AM Beijing time. The overseas DevPack is almost always in stock. And with annual billing plus discount codes, the effective overseas price can drop significantly.

GLM gotchas: peak-hour multipliers and the Lite model trap

Three things catch people off guard with Zhipu GLM:

• GLM-5 and GLM-5.1 are restricted to Pro tier and above. Lite subscribers only get GLM-4.7. If you bought Lite expecting the flagship model, you will be disappointed.
• Peak-hour quota multiplier: from 14:00 to 18:00 UTC+8, GLM-5 consumes 3x quota per request. Off-peak is 2x. You think you have 400 prompts, but during peak hours you actually get about 133. A temporary off-peak promotion (1x multiplier) was running through late April 2026, but this is not guaranteed to last.
• Domestic Max tier has severe availability limits. Only about 20% of daily quota opens at 10:00 AM Beijing time, and it frequently sells out within minutes. Set your alarm or buy overseas.

Kimi (Moonshot AI): not just a coding tool — an AI work style

Like GLM, domestic pricing is far below overseas. Domestic Moderato at ¥79 is roughly half the price of the overseas equivalent at $19 (about ¥152). The overseas-only Vivace tier ($199/month) unlocks 4 concurrent agents plus 8 sub-agent swarms — there is no domestic equivalent at that scale.

But here is what makes Kimi different from every other provider on this list: the ¥39 Andante tier includes 10 Agent sessions per month, 10 deep research sessions, 10 PPT generations, plus Kimi Code programming quota. For people who are not full-time programmers but occasionally need to write code, do research, or create presentations, this “everything bundle” is more practical than any pure coding plan.

Kimi is not selling a coding tool. It is selling an AI work style. That positioning separates it from every other provider. My team has product managers who use Kimi all day — and they compete with the engineers for Agent quota.

Official screenshot

Kimi K2.5 is best introduced from the technical blog, not from social summaries

The public K2.5 tech blog already combines the multimodal upgrade story, benchmark tables, and Agent Swarm explanation in one official page.

• Best official visual for the K2 to K2.5 family transition.
• Useful for readers who need proof that Agent Swarm and MoonViT are part of the public story.

Source: Official Kimi K2.5 tech blog.

MiniMax: the absolute cheapest entry — but the domestic version is quota-capped

On the model side, M2.7 scored 56.22% on SWE-Pro and 57.0% on Terminal-Bench 2.0. The older M2.5 was even more impressive: SWE-bench Verified 75.8%-80.2% (official leaderboard 75.80%, self-reported maximum 80.2%), placing it among the very best open-source models.

What makes MiniMax unique is full-modal bundling — images, voice, music, and video are all included in the same plan at no extra cost. Nobody else offers that. ¥29/month does not just buy you a coding tool; it buys you an AI multimedia studio.

Official screenshot

MiniMax clearly positions Token Plan as the current subscription route

The official Token Plan overview is the best first stop for public-facing articles because it explains the route before readers ever hit a pricing table.

• Useful for queries around MiniMax Token Plan, MiniMax Coding Plan, and MiniMax subscription.
• Helps clarify that Token Plan is the current public route readers should treat as primary.

Source: MiniMax Token Plan overview.

Official screenshot

MiniMax publishes a strong public pricing table for monthly and Highspeed tiers

This pricing-table view is one of the best official screenshots in the category because readers can verify standard and Highspeed tiers directly from the source page.

• Shows the monthly standard tiers and the Highspeed plan table in one view.
• A good visual checkpoint before repeating plan prices or 5-hour request limits in an article.

Source: MiniMax Token Plan pricing.

Xiaomi MiMo: finally, a plan without the 5-hour window trap

The overseas starting price of approximately $5/month may be the cheapest coding plan entry point anywhere in the world. Note that Xiaomi has said “global launch” but has not formally published overseas pricing — treat the USD numbers as estimates.

MiMo-V2-Pro is the flagship model: over 1 trillion total parameters (MoE architecture, 42B activated), with a 1M context window. Xiaomi claims coding performance surpasses Claude 4.6 Sonnet and approaches Opus 4.6, though independent third-party benchmark verification is still pending. PAYG API costs are reportedly about 20% of competing models.

I know several engineers who do agent debugging for a living, and they love the no-window model. Their work pattern is “sit still all day, then suddenly get three hours of intense inspiration.” MiMo Credits match that rhythm perfectly.

Good product design is not about giving you more — it is about not setting limits. MiMo understood this.

Official screenshot

MiMo-V2-Pro now has a strong official release page with buyer-relevant claims

The Xiaomi release note is the best page to anchor the 1T / 42B / 1M context story before you move into the pricing and integration docs.

• Useful for replacing older beta-era summaries with official product language.
• Pairs naturally with the pricing and tools overview pages for route clarity.

Source: Official MiMo-V2-Pro release note.

Official image

Xiaomi publishes the Artificial Analysis ranking image directly on the official MiMo-V2-Pro page

The buyer-facing MiMo page does not only describe the model in prose. It also exposes the ranking visual that Xiaomi uses to support the “8th worldwide, 2nd among Chinese LLMs” positioning.

• Useful when readers want a traceable official image instead of a copied leaderboard screenshot from social posts.
• Works well alongside the pricing page because it keeps performance proof and route proof on official Xiaomi surfaces.

Source: Official MiMo-V2-Pro page.

StepFun (Jieyu Xingchen): the one where overseas is actually cheaper

”Promotional period” is the countdown timer on every discount. You think you are gaming the system. They are building your habits.

• All prices are currently labeled “promotional period pricing” — StepFun has not promised these rates are permanent.
• Billing is prompt-based, where 1 prompt equals roughly 15-20 model calls. The prompt count looks low, but each prompt does more work.
• Step Plan is tightly coupled to the OpenClaw ecosystem. If you are a heavy OpenClaw user, the value proposition is strong. What happens after the promotional period ends is an open question.

Official screenshot

Step 3.5 Flash exposes architecture and benchmark rows directly on the official repo page

For public writing, the GitHub page is one of the best official assets because it makes the MoE architecture, benchmark table, and open-weight route visible without guesswork.

• Strong image for explaining why Step 3.5 Flash is framed around speed and decoding efficiency.
• Useful when readers want to verify that the model is an Apache 2.0 open release.

Source: Official Step 3.5 Flash GitHub repository.

Alibaba Bailian: the ¥40 entry tier vanished overnight

Overseas Pro at $50 (about ¥400) is double the domestic ¥200 price. Like GLM and Kimi, domestic users get a substantial price advantage.

But Pro has something nobody else offers: multi-model aggregation. One subscription gives you access to Qwen3-Coder-Next, Kimi-K2.5, GLM-5, and MiniMax-M2.5. It is like having one SIM card that works on every carrier network simultaneously. Plus it supports both OpenAI-compatible and Anthropic-compatible endpoints.

When every other provider is selling their own model, Bailian chose a different path: selling a model supermarket. Whether that works long-term depends on whether the shelves stay stocked with top-tier models.

• Critical technical gotcha: Bailian requires a dedicated Coding Plan API key (prefixed sk-sp-...). Using a regular API key will trigger pay-as-you-go billing, which can cost 5x more than the plan. I have seen someone burn through over ¥1,000 in a single month because they used the wrong key.
• The entry barrier jumping from ¥40 to ¥200 is genuinely hostile to students and newcomers.
• Multi-model access in one plan is excellent for anyone doing comparative model evaluation — no more juggling five different accounts and API keys.

Official screenshot

Qwen3.6-Plus already has route-specific public pricing on the Model Studio side

The Alibaba pricing page is the safest public surface for route-aware Qwen3.6-Plus billing because it distinguishes mainland and international rows directly on the official page.

• Useful when articles need one source-backed image for regional pricing differences.
• Pairs well with benchmark claims from the Qwen 3.6 release page.

Source: Alibaba Cloud Model Studio pricing.

Volcengine Ark (ByteDance): the design-to-code frontend specialist

Doubao Seed 2.0 Code deserves special attention. It swept all five gold medals in ICPC programming competitions. The community has positioned it as “the visual-driven frontend weapon” — if your workflow is “design mockup to working code,” this is currently the best value option.

Volcengine also fell into the first-purchase discount trap: Lite was ¥9.9 and Pro was ¥49.9 for your first purchase, but from March 13 onward, those became daily 10:30 AM limited flash sales. Yes, you read that right — you now set an alarm to抢 (snatch) token plans. Concert tickets, Moutai liquor, and now AI subscriptions. This is what 2026 looks like for programmers.

Pro subscriptions also include ArkClaw Lite. Users already in the ByteDance ecosystem (TRAE, Doubao App) will find the integration natural. As with Bailian, make sure you use the dedicated Code Plan URL — traffic through the generic API endpoint is not counted toward your plan quota.

Official screenshot

The official Seed2.0 page is the cleanest public source for the model-family story

ByteDance's Seed site makes the Pro, Lite, and Mini lineup, benchmark breadth, and Ark access links visible in one place. That is stronger than building the article around unofficial pricing screenshots.

• Best official visual for the Seed2.0 family overview.
• Useful for replacing reseller-style descriptions with Seed's own model framing.

Source: Official Seed2.0 model page.

Benchmark comparison: the models behind the plans

Choose your model before you choose your plan. Choose your model based on what you actually do. Comprehensive bug-fixing? Kimi K2.5 or MiniMax M2.5. The hardest engineering challenges? GLM-5.1. Frontend work from designs? Doubao Seed 2.0 Code.

Domestic vs overseas: why mainland users get better deals (and when they do not)

Why is domestic so much cheaper for most providers? Because the Chinese market is a seven-way knife fight for users. Seven providers in one pool, all slashing prices. The overseas market faces Claude, GPT, and Gemini — a completely different pricing logic where you do not need to win on price, you need to prove “I am not worse than Claude.”

The cheap prices you get domestically exist not because the service is truly cheap, but because someone is fighting a price war on your behalf. That war will not last forever. Bailian Lite is already gone. Volcengine first-purchase discounts are now flash sales. GLM has raised prices twice.

Those cheap tiers are disappearing one by one.

Will tokens keep getting more expensive?

Yes. And this process has barely started.

Late 2024 through mid-2025 was the “subsidize to acquire users” phase. Per-million-token prices dropped from tens of yuan to fractions. Cheap enough that you did not bother checking usage. From late 2025 onward, the screws tightened: first-purchase discounts eliminated, low tiers discontinued, flash sales replacing open stock. In 2026, we entered the “fine-grained segmentation” phase: MiniMax launched Highspeed tiers from ¥98 to ¥899, Kimi introduced the ¥559 Allegro tier.

Tokens are being tiered exactly like mobile data plans were. Daily passes, monthly bundles, unlimited-with-throttling — every playbook from the telecom era is being replayed in AI with remarkable precision.

The root cause is compute cost. Training a hundred-billion-parameter model runs into the hundreds of millions of dollars, and inference is not cheap either. The low prices were always venture-funded customer acquisition, and investor patience is finite.

Will the market consolidate down to a few suppliers? Very likely. Look at cloud computing today — AWS, Azure, and Alibaba Cloud in a three-way stand. The AI token supply chain will probably follow the same pattern. Aggregation platforms like Bailian and Volcengine are becoming the “virtual operators” of the AI era — they do not build base stations, they package several providers' signals and sell them to you.

• Budget at list price, not promo price. First-purchase discounts only apply once. Flash sales are not guaranteed. Promotional periods end without warning.
• The number you can sustainably pay every month for the foreseeable future — that is your real cost.
• Consolidation is coming. The seven providers today may become three or four within a year or two.

If I were buying today: budget-tier recommendations

¥29/month will not get you the strongest model. ¥200/month will. Which is the better deal? The answer does not depend on the price — it depends on your use case. If you only code for 30 minutes a day, ¥29 is enough. If your livelihood depends on AI-assisted programming, ¥200 is the frugal choice.

Six buying rules to remember

• Pick the model first, then find the plan. GLM-5.1 for hard engineering, Kimi K2.5 for comprehensive code repair, Doubao Seed 2.0 for frontend from designs, MiniMax for the cheapest entry with full-modal support.
• Check which route you are actually buying. A membership is not a token plan is not a coding plan is not a PAYG API. These are fundamentally different products that happen to live on the same provider website.
• Use the correct API key. Bailian requires sk-sp-... prefixed keys. GLM has separate DevPack and API routes. Using the wrong key or endpoint can cost you 5x more than the plan price.
• Watch for peak-hour multipliers. GLM-5 uses 3x quota during peak hours (14:00-18:00 UTC+8). Your 400-prompt allowance drops to 133 effective prompts when you need them most.
• Budget at regular price, not promo price. First-purchase discounts expire. Flash sales are unreliable. Promotional periods end. The sustainable monthly cost is what matters for planning.
• If you are overseas, do not just convert CNY prices at face value. Check whether the overseas tier gives different quotas (MiniMax gives 2.5x more), has different purchase limits (GLM overseas has no stock caps), or is actually cheaper (StepFun Flash Plus).