Coding Plan vs Token Plan vs API Aggregator in November 2026: What the Alibaba and Baidu Restructurings Changed
Two domestic vendors restructured their plan tiers in the same week, and both moves change the arithmetic behind the subscription-versus-API decision. Alibaba Cloud's November 9 notice raised credit allowances by 74 to 78 percent on the three lower personal-edition tiers while holding their monthly fees flat, then cut the Pro tier price 37 percent from CNY 499 to CNY 312. Baidu Qianfan inserted a Standard tier between Lite and Pro during Double 11, adding a CNY 120 list price that did not exist a week ago. Priced from the vendors' own pages as of November 11, 2026: IDE coding subscriptions (Cursor Plus $8 to Ultra $200, GitHub Copilot Pro $10 to Max $100, Devin Desktop Pro $20, OpenCode Go $10), Chinese token and credit plans (Alibaba Bailian CNY 39 to 312, Baidu Qianfan CNY 9.9 to 600, Tencent Hy3, Volcengine Agent Plan, Xiaomi MiMo, MiniMax, Kimi Code CNY 49 to 699, Zhipu GLM Coding CNY 118 to 1078), and API aggregators. Includes the credit-to-token conversion math under the new allowances, which discounts expire this month, and a decision rule mapping each billing model to the workload it actually fits.
Choosing a billing model before choosing a model sounds like an unnecessary abstraction, but in November 2026 it is the decision that dominates the bill. Two domestic vendors moved inside the same week. Alibaba Cloud's official notice of November 9 raised credit allowances by 74 to 78 percent on the three lower personal-edition tiers while holding their monthly fees flat, and cut the Pro tier from CNY 499 to CNY 312 while still raising its quota. Baidu Qianfan inserted a Standard tier at CNY 120 list price between Lite and Pro during its Double 11 window, which closed a gap that had forced users to jump from CNY 40 straight to CNY 200. Both moves make subscription units cheaper, and both make the break-even point against pay-as-you-go API pricing move further out. The three billing models still behave fundamentally differently underneath, and the practical question is which one matches how your load is actually shaped.
What actually changed this week
Alibaba Cloud's notice of November 9 restructures all four Token Plan personal-edition tiers. Three hold their monthly fee while their credit allowance jumps: Lite from 11,500 to 20,000 credits, Essential from 25,500 to 45,000, Standard from 45,000 to 80,000. Pro is the outlier, with both the list price and the on-site promotional price down 37 percent from CNY 499 to CNY 312 while the quota still rises from 180,000 to 200,000 credits. Existing subscribers are topped up to the new allowance inside the current billing cycle at no extra cost, and Pro's new price applies from the next renewal. Baidu Qianfan's move is narrower but fills a real gap. Its Double 11 window, running from November 6 to November 23, adds a Standard tier at CNY 120 list price for 22,000 points, sold at CNY 49.9 for new subscribers for 23,760 points, with renewals at CNY 69.9. Before this there was no tier between Lite at CNY 40 and Pro at CNY 200, so a user needing more than Lite had to buy five times the price. No API vendor changed a price this week.
Coding subscriptions: you are buying a workflow, not a token
The international coding subscriptions did not move this week, and their current structure is worth restating because the tiers are not a price ladder over one product. Cursor runs Free at zero, Plus at USD 8, Premium at USD 30, Super at USD 100 and Ultra at USD 200, with usage multipliers of 4x, 15x and 50x against Plus, and a material restriction that the USD 8 Plus tier only exposes Grok models while third-party models bill separately at API list rates. GitHub Copilot offers Pro at USD 10, Pro+ at USD 39 and Max at USD 100, running on AI Credits where code completion stays unlimited and the Pro tier ships USD 15 of allowance. Devin Desktop Pro at USD 20, the entity Windsurf became after Cognition acquired it, allows buying extra usage at API pricing. OpenCode Go at USD 10 quotes its allowance per model rather than as a flat token pool. The point that matters for this week's comparison is that none of these scale with the token volume you actually consume, so a subscription that covers your interactive work does not cover the batch job you run overnight.
The new allowance math: what a credit actually costs now
Converting each tier to cost per credit is what makes the two restructurings comparable, and the ranking changed. Under the old Alibaba allowances, Pro at CNY 499 for 180,000 credits was the cheapest unit in the market at CNY 0.0028 per credit, ahead of Lite at CNY 0.0034. After the restructuring Lite falls to CNY 0.0020 per credit at CNY 39 for 20,000 credits, Standard sits at CNY 0.0017 at CNY 139 for 80,000 credits, and Pro holds CNY 0.0016 at CNY 312 for 200,000 credits. Every Alibaba tier now undercuts its own previous best, and the entry tier became the most improved. Baidu's new Standard tier lands at CNY 0.0055 per point at the CNY 120 list price, between Lite at CNY 0.0035 and Pro at CNY 0.0044, while the new-subscriber promotion on it drops the effective rate to CNY 0.0021. The caveat is unchanged and decides everything: these rates only hold if you consume the full allowance, and the allowance resets rather than accumulating.
The API side and the aggregator: no price moved, but the comparison moved
Not one pay-as-you-go price changed this week, which means the API floor is exactly where it was: Qwen3.8-Flash and GLM-5.3-Flash at CNY 0.80 input per million tokens, MiniMax M3.1 Flash at CNY 1 and DeepSeek V4.1 Flash at CNY 1 off-peak with a USD 0.003 cache-hit rate, all carrying 1M context. On the dollar side GPT-6 Luna and Claude Haiku 5.5 both sit at USD 0.10 input and USD 0.50 output, while GPT-6 Astra caps the output side at USD 50. Aggregators did not discount either, and the honest summary of their proposition in November 2026 is unchanged: TokenRhythm's cheapest 1M-context routes are the vendors' own list prices, and what you buy is one key across 30 plus models, dual protocol compatibility, semantic routing, CNY 68 in signup credits, no card required, and upstream failover without client-side rewrites. The arithmetic shift is that subscription unit costs fell while API prices held, so the crossover point moved further toward the API side for anyone whose load is heavy and predictable.
One rule for choosing, and which discounts to act on first
Normalizing to cost per million tokens or per credit remains the only comparison that survives a real invoice, and this week's numbers give a clearer answer than usual. If your load can be described in requests per day you are in the subscription market, and the November numbers make the entry tiers unusually attractive: Baidu Mini at CNY 4.9 for the first month, Alibaba Lite at CNY 19.50 for new users during Double 11, Cursor Plus at USD 8. If it can be described in tokens per month you are in the API market, where the CNY 0.80 floor and the USD 0.003 cache rate remain unbeaten. If you need more than one vendor behind one key you are in the aggregator market and should read the per-model catalog rather than the headline. On timing, two promotions are about to expire: Volcengine's Kimi K3.1 Pro 50 percent coefficient discount ended November 9 and its Coding Plan 25 percent promotion ended November 8, while Tencent Cloud's TokenHub activity giving DeepSeek V4 Pro 30 percent off and V4.1 Flash 60 percent off runs to November 16 at 12:00 and Baidu's Double 11 window closes November 23.
Frequently Asked Questions
What exactly did Alibaba Cloud change in Token Plan on November 9?
All four personal-edition tiers were restructured. Lite keeps CNY 39 per month and its allowance rises from 11,500 to 20,000 credits, an increase of 8,500 credits. Essential keeps CNY 79 and rises from 25,500 to 45,000 credits, adding 19,500. Standard keeps CNY 139 and rises from 45,000 to 80,000 credits, adding 35,000. Pro is the exception: both the CNY 499 list price and the on-site promotional price drop 37 percent to CNY 312 per month, while the allowance still rises from 180,000 to 200,000 credits. The four tiers take effect for existing subscribers from November 9 at 20:00 Beijing time, and subscribers are topped up to the new allowance within the current billing cycle with already-consumed credits left unchanged. Pro users keep the old price for the current cycle and the new price applies on the next renewal, and if the current cycle quota is already exhausted the difference is credited within the same cycle.
What is the new Baidu Qianfan Standard tier and who is it for?
Baidu added a Standard tier during its Double 11 window, running from November 6 at 00:00 to November 23 at 23:59. The list price is CNY 120 per month for 22,000 points, and during the window new subscribers pay CNY 49.9 per month for 23,760 points, which is 8 percent more than the list quota. Renewals run at CNY 69.9 per month, 30.1 percent below list. The tier exists to fill the gap between Lite at CNY 40 for 11,500 points and Pro at CNY 200 for 45,000 points, which previously forced anyone needing more than Lite to buy five times the price. It is aimed at developers whose usage outgrew Lite but did not justify Pro, and the off-peak discount of 20 percent from 22:00 to 08:00 on DeepSeek-V4.1-Flash, DeepSeek-V4-Pro-0813 and GLM-5.3 applies to it as well, running until November 30.
After the restructuring, which Chinese subscription tier has the cheapest unit cost?
Alibaba Cloud Pro is the cheapest unit rate at CNY 0.0016 per credit, based on CNY 312 for 200,000 credits, followed by Standard at CNY 0.0017 from CNY 139 for 80,000 and Lite at CNY 0.0020 from CNY 39 for 20,000. Before the restructuring the ranking was Pro at CNY 0.0028, Lite at CNY 0.0034 and Standard at CNY 0.0034, so every tier now undercuts the previous best single rate. Baidu's tiers are higher: Lite at CNY 0.0035 per point, Pro at CNY 0.0044, Max at CNY 0.0036, and the new Standard at CNY 0.0055 at list price, though its CNY 49.9 new-subscriber promotion brings the effective rate down to CNY 0.0021. The decisive caveat is that a unit rate is only real when the allowance is fully consumed, and allowances reset at the cycle boundary rather than accumulating.
Do I now have a cheaper option than paying per token for a production API workload?
For most production workloads, no, and this week's restructuring did not change that. Subscription allowances are denominated in credits whose conversion to tokens depends on which model you call and with what cache hit rate, so the effective per-million-token cost of a subscription is not a single number. More importantly, the allowance resets. A monthly plan sized for a spiky workload loses whatever you did not consume, while pay-as-you-go billing is metered exactly. The restructuring moved the break-even point in a specific direction: subscription unit costs fell by roughly 35 to 45 percent on Alibaba's lower tiers while API list prices held, so if you consume your allowance fully every month, a subscription now looks better than it did a week ago. If your usage is variable or you run multiple models with different cache behaviour, per-token billing with cache hits and Batch calls still wins.
Which discount should I act on before it expires?
Two Volcengine promotions have already ended, on November 8 and November 9, so treat those as gone. The most urgent remaining one is Tencent Cloud's TokenHub Token Plan activity, which gives DeepSeek V4 Pro 30 percent off and DeepSeek V4.1 Flash 60 percent off until November 16 at 12:00 Beijing time, with idle and peak hours priced separately. Baidu's Double 11 window closes November 23 at 23:59, and it covers the new Standard tier, the five-fold first-purchase offer from CNY 4.9 and the Mini limited-time offer, while the separate off-peak discount runs to November 30. Alibaba's Double 11 new-user prices of CNY 19.50 for Lite and CNY 39.50 for Essential expire November 30. The discipline that survives every week: plan against list prices and treat promotions as upside, because Zhipu already runs GLM-5.3-Flash with no published end date on its discount and can stop it without warning.
What is the full domestic plan ladder with prices and allowances after this week's changes?
Alibaba Cloud Bailian personal tiers after November 9: Lite CNY 39 for 20,000 credits, Essential CNY 79 for 45,000, Standard CNY 139 for 80,000 with Harness benefits, Pro CNY 312 for 200,000 credits; team seats from CNY 150. Baidu Qianfan personal tiers: Mini CNY 9.9 for 1,400 points, Lite CNY 40 for 11,500, the new Standard CNY 120 for 22,000, Pro CNY 200 for 45,000 and Max CNY 600 for 165,000, with 20 percent off-peak from 22:00 to 08:00 until November 30. Volcengine Ark: Agent Plan CNY 40 for 2,000 AFP per 5 hours and 20,000 AFP per month, first two months CNY 9.9; Coding Plan Lite CNY 40 for 148 requests per 5 hours, 320 million tokens per month and 256K context. Tencent Hy3 Token Plan CNY 28 for 560 points. Xiaomi MiMo Token Plan Lite CNY 34 continuous monthly, 1M context, no 5-hour limit, 0.8x consumption between midnight and 08:00. MiniMax Starter CNY 29 and Plus CNY 49, where Plus covers roughly 600 million M3 tokens a month. Kimi Code CNY 49, 99, 199 and 699 for 1x, 4x, 20x and 60x weekly allowance. Zhipu GLM Coding Lite CNY 118, Pro CNY 538 and Max CNY 1078 for 10,000, 60,000 and 140,000 credits a week. International equivalents: Cursor Plus $8, Premium $30, Super $100, Ultra $200; GitHub Copilot Pro $10 for $15 of allowance, Pro+ $39 and Max $100; Devin Desktop Pro $20; OpenCode Go $10, Go Plus $40 and Go Teams $20 per member; Z.ai GLM Plan $18, $80 and $168.
What is the difference between a Token Plan and a Coding Plan?
The billing unit and the acceptable usage. A coding plan meters requests or sessions for interactive IDE work, which is why Cursor Plus at USD 8 only exposes Grok models and why Devin Desktop caps concurrent sessions at 10. A token plan meters credits or points against a published monthly allowance and can be pointed at any model the vendor hosts, which is why Alibaba quotes 20,000 to 200,000 credits per month after this week's restructuring and Baidu quotes 1,400 to 165,000 points. A token plan is what a production service should run on, a coding plan is what a human should use, and the credentials are not interchangeable: Token Plan issues a dedicated key such as sk-sp-xxxxx with its own base URL that cannot pay for pay-as-you-go requests.
Want more AI pricing comparisons?
Check our price changelog for daily updates on AI model pricing, new releases, and deprecations.