1. Token Intelligence Brief

As we said at the close of last issue's individual-focused roundup, this issue returns to the enterprise and organizational perspective. The vendors' price panels happen to be in a rare quiet spell — DeepSeek's two tiers have gone nine consecutive issues without a move, Zhipu's GLM-5.3 has held flat for nine straight issues, and Kimi K3.1 Pro has not changed in four — which makes this the right moment for something we normally cannot do: put "how you buy" on the table and compare it head-to-head. Which unit price is lower is plain at a glance, but the price gap between channels for the same model, and the bill gap between plans for the same volume of business, is a column many procurement sheets still do not have. First the panel, then the comparison.

A. Official Pricing on Domestic Platforms (RMB per 1M tokens)

Platform Model Input Output Notes
DeepSeek deepseek-flash Off-peak 1.00 / Peak 2.00 Off-peak 4.00 / Peak 8.00 Cache hit from 0.02; concurrency 2500
DeepSeek deepseek-v4-pro Off-peak 4.50 / Peak 9.00 Off-peak 13.50 / Peak 27.00 Cache hit from 0.15; concurrency 500
Zhipu GLM-5.3 (flagship) 8.00 28.00 Cache hit RMB 2; Batch API at 50% off; 1M context
Zhipu GLM-5.3-Flash 0.80 2.80 Limited-time 50% off: input 0.4 / output 1.4
Kimi (Moonshot AI) Kimi-K3.1-Pro 13.00 (cache hit 1.30) 68.00 1M context; Batch at 60% of standard price
Kimi (Moonshot AI) Kimi-K2.7-code 6.50 (cache hit 1.30) 27.00 Closed to newly registered users from October 31
Alibaba Cloud Bailian Qwen3.8-Max 12.00 36.00 Batch at half price; free quota of 1M tokens
Alibaba Cloud Bailian Qwen3.8-Flash 0.80 2.70 New users get 1M tokens free (valid 90 days)
ByteDance Volcano Ark Doubao Seed 2.1-Pro 6.00 30.00 Cache hit 1.2; low-priority tier 3/15

DeepSeek time-of-day rules: 9:00-12:00 and 14:00-18:00 on weekdays are peak; all other hours (including all day on weekends and statutory holidays) are billed at off-peak rates, roughly half of peak. Kimi states officially that its API is pay-as-you-go with no subscription plans; Kimi's file-related APIs (document extraction/storage) are free for a limited time. Zhipu's GLM-4.7-Flash is free across input, output, and cache (200K context) — still the usual starting point for zero-cost pilots.

B. Channel Price Comparison (New This Issue: Two Ways to Buy the Same Model)

Model Official direct (RMB) Aggregator platform price (RMB basis) Key differences
DeepSeek V4-Pro Off-peak 4.5/13.5, Peak 9/27 12/24 (flat rate) The aggregator has not followed the official price cut; its input price is 2.7x the official off-peak
DeepSeek V4.1-Flash Off-peak 1/4, Peak 2/8 Off-peak matches the official price, but off-peak counts only 2-8 The aggregator bills evenings and weekends back at full price
GLM-5.3 8/28 8/28 Identical to Zhipu's official price

The value of third-party aggregator platforms (one-stop access to many models) lies in "signing fewer contracts and calling many models through one interface", but their prices do not always track the official ones: some tiers match official pricing while others run several times above the official off-peak. The table above compares currently listed prices; before procuring, defer to each platform's real-time page.

C. Subscription Plans at a Glance (For Enterprises & Developers)

Platform Plan Price (currently listed) Key benefits
Zhipu GLM Coding Plan Lite RMB 94.4/month on continuous billing (list price 118) 10,000 credits per week, works with Claude Code and 20+ coding tools
Zhipu GLM Coding Plan Pro / Max RMB 430.4 / 862.4 (20% off on continuous billing) 6x/14x the Lite quota
Zhipu Team edition standard seat About RMB 538.2/month and up after the 10% annual-billing discount 66,000 credits per week; unified management of organization seats and permissions
Zhipu Team edition premium seat About RMB 1,078.2/month and up after the 10% annual-billing discount 155,000 credits per week; priority resource guarantees during peak periods
Anthropic Claude Pro About RMB 121/month billed annually ($17), about RMB 142 paid monthly ($20) No API credits
Anthropic Claude Max $100/month Includes $100/month in API credits
Anthropic Claude Ultra $200/month Includes $200/month in API credits
Anthropic Claude Team $20/seat/month + usage at API prices Team-shared API credits, up to $500 a month

The API credits bundled with subscriptions are a "deduction channel", not an API price cut: real-money usage still burns credits at the API list price, and once the credits run out you pay the full rate. Kimi is pay-as-you-go with no subscription tier; its Batch API (60% of standard price) is the only discount channel on its official books.

D. Free and Low-Cost Roundup (Enterprise View)

  • Zhipu GLM-4.7-Flash (200K context): input, output, and cache all free; the vision model GLM-4.6V-Flash is free as well. A zero-cost starting point for prototype validation — the free tier carries concurrency limits, so switch to a paid tier before scaling into production.
  • Alibaba Cloud Bailian: new users get 1M tokens free (90 days); the Night Plan overnight 40%-of-list limited-time campaign is still running (jobs start after 22:00; the applicable scope is subject to the campaign page's real-time listings).
  • Tencent Hunyuan: a 1M-token resource pack for new users, valid for 1 year.
  • China National Supercomputing Internet: the RMB 0.1 special-offer resource pack currently contains 50M tokens (about RMB 0.002 per 1M tokens) — the lowest publicly visible unit-price anchor on the open web; log in to the platform to verify detail billing before scaling up.
  • Reminder: DeepSeek's Search/Fetch web-connected tools are billed per use from 0:00 on November 1 (RMB 5 per 1,000 successful requests each); last issue laid out three usage-tier estimates. November is the month it lands on its first complete monthly bill, and teams doing month-end reconciliation should remember to list it as a separate line item.

Data note: the price data in this article was obtained from publicly available data on the official platforms and is the latest data at press time.


2. Focus of This Issue: Same Model, Different Channel — a 2.5x Difference in the Bill

Manager Wang from a friend's company came by last week with a quotation sheet to check: they were hooking an internal system up to a flagship model for document processing; the selection meeting had settled on DeepSeek V4-Pro, the tech department had opened the account through an aggregator platform, and a month of running had cost over RMB 1,300 in model fees. His first question was: "Is this price inflated?"

Put the two channels' list prices side by side and the answer appears. The same DeepSeek V4-Pro: official direct is time-of-day pricing — off-peak input RMB 4.5, output RMB 13.5 per 1M tokens, peak RMB 9/27; the aggregator is a flat rate — input RMB 12, output RMB 24. On input price alone, RMB 12 against RMB 4.5, the aggregator costs 2.7x more; the output unit price is nearly 80% higher too. Manager Wang's internal document processing runs about 95M tokens of input and 8M tokens of output a month and can be scheduled into evenings and weekends — hours that count as off-peak around the clock under the official rules. For the same work, the aggregator's flat-rate bill comes to about RMB 1,332 a month, while running official direct at off-peak costs only about RMB 536 — a gap of nearly 2.5x; even billed entirely at peak rates, official direct comes to about RMB 1,071, still nearly 20% cheaper than the aggregator.

First, the nature of this: nobody here is "mislabeling prices". The aggregator settles at a flat rate; the official channel settles by "what hour you use it" — two different billing logics. The root of the gap is that DeepSeek built time-of-day discounts into its official unit-price table — the off-peak price is only half of peak, and the aggregator channel has not followed that half. The industry runs other forms of the same play: Alibaba has made overnight 40% of list a limited-time campaign, and ByteDance has priced its low-priority tier at about half price — "trading time or priority for money" takes a different shape at each vendor, but the discount layer held by official channels is generally thicker than the one held by reseller channels.

The price-comparison advice for enterprises comes in three points:

First, run your main models on official direct. For the two or three models with the heaviest usage, open accounts directly on the official open platforms — only there can you reach the official discount layers: time-of-day pricing, Batch, cache. The aggregator settles at a flat rate, and what the premium buys is the convenience of "managing fewer accounts".

Second, treat the aggregator as a test plot, not the main field. During selection, trying five or six vendors' models from one console, the aggregator genuinely saves effort; but once results are in and the main models are chosen, migrate production usage back to official direct — usually something a single reconciliation statement makes plain.

Third, write "time-slot flexibility" into the selection criteria. When reviewing models for procurement, add one line beyond unit price: does this vendor have off-peak pricing, Batch pricing, cache pricing? Vendors with all three (DeepSeek, Zhipu, and Kimi, for example) do twice the work on the same budget.

The bottom line: the first row of a price-comparison sheet should not be "which vendor is cheaper" but "which channel of the same vendor is cheaper" — one line like that is enough to save buyers like Manager Wang the cost of a team-building dinner on next month's bill.


3. Token Ledger: Three Enterprise Accounts, Each with Two Endings

Everything below is an estimate: scenarios are built from typical business volumes, actual costs float with the usage mix, and international prices are converted at 1 USD ≈ RMB 7.1. The three accounts map to the three most common enterprise consumption patterns: "asks a lot", "reads a lot", and "runs late". Last issue's three accounts covered "reading approach and architecture" (RAG, segmentation, cache); this issue's three cover "how you buy": channel, tier, and billing terms.

Scenario 1: Online Customer-Service Bot (Mid-Size) — from RMB 74 to RMB 3,120

Setup: an SMB online customer service whose combined monthly conversation input is about 200M tokens (about 120K questions, each carrying knowledge-base context) and whose output is 20M tokens; hit-type questions account for 60% and run at cache-hit prices.

Plan Estimated monthly cost Notes
GLM-5.3-Flash 50% off (0.4/1.4, cache 0.115) about RMB 74 the limited-time 50%-off basis, maximum-compression tier
DeepSeek flash off-peak (1/4, cache 0.02) about RMB 162 customer-service consultations skew heavily to nighttime, a natural fit for off-peak pricing
MiniMax-M3.1-Flash (1/4, cache 0.10) about RMB 172 1M context is rare at this price point
DeepSeek flash peak (2/8, cache 0.04) about RMB 325 the basis when consultations concentrate in daytime
Doubao Seed 2.1-Pro (6/30, cache 1.2) about RMB 1,224 a major vendor's workhorse tier
GLM-5.3 (8/28, cache 2) about RMB 1,440 straight to the flagship
Kimi K3.1 Pro (13/68, cache 1.3) about RMB 2,556 the 1M-capability tier, which customer service cannot fully use
Qwen3.8-Max (12/36) about RMB 3,120 the flagship ceiling for reference

The same customer service runs from RMB 74 to RMB 3,120 — a gap of roughly 42x. Issue 2 ran the numbers for a large-scale service averaging 10,000 conversations a day (monthly bills from RMB 740 to 36,000); this issue shrinks the scale to SMB size, and the conclusion holds as-is: customer-service Q&A does not need flagship reasoning. Lightweight tiers plus cache pin the bill in the RMB-hundreds range; in this scenario the flagship tier means paying big money for reasoning depth you cannot use. What deserves attention instead is the "time slot": consultation volume keeps flowing at night, and DeepSeek's off-peak pricing counts nights and all day on weekends — of all business types, customer service is the one that should be harvesting this dividend most.

Scenario 2: Batch Bid/Contract Review (100 Documents a Month) — from RMB 18 to RMB 540

Setup: each document is about 400K tokens (a bid of roughly 600K characters), output is a key-point extract of 3,000 tokens, 100 documents a month. The two reading approaches continue to follow issue 6's conclusions: element-extraction reviews can be read in segments (counted at a 10% segment-overlap overhead), while work needing cross-clause reconciliation goes to 1M whole-read.

Plan Estimated monthly cost Reading approach
GLM-5.3-Flash 50% off, segmented (0.4/1.4) about RMB 18 segmented read
DeepSeek flash off-peak, segmented (1/4) about RMB 45 segmented read, run in the evening
DeepSeek flash peak, segmented (2/8) about RMB 91 segmented read, run in daytime
GLM-5.3 Batch whole-read (4/14, 50% off) about RMB 164 1M whole-read, async at half price
Kimi K3.1 Pro Batch whole-read (7.8/40.8, 60% of list) about RMB 324 1M whole-read, async at 60%
GLM-5.3 pay-as-you-go whole-read (8/28) about RMB 328 1M whole-read, real-time
Kimi K3.1 Pro pay-as-you-go whole-read (13/68) about RMB 540 1M whole-read, real-time

The same batch of bids runs from RMB 18 to RMB 540 — a gap of roughly 30x. What this issue wants you to give a second look is the middle two rows: for the same "1M whole-read, never cut apart", going through the Batch channel (Zhipu at 50% off, Kimi at 60% of list — batch accepts "the results arrive a bit later", and the bill drops straight from RMB 328 to 164 and from RMB 540 to 324). Bid review never lacked those one or two hours anyway; this is nearly half saved by touching nothing but the submission method — not the model, not the reading approach.

Scenario 3: Night Batch Production Line (Data Processing/Content Production) — from RMB 227 to RMB 2,904

Setup: a production line with monthly input of about 264M tokens and output of about 132M tokens, jobs queueable with no real-time return required. Issue 6 ran this same scenario; this issue fills in the rest of the "official-channel billing-term discounts":

Plan Estimated monthly cost Notes
Alibaba Qwen3.8-Flash overnight 40% of list (about 0.32/1.08) about RMB 227 Night Plan basis, runs after 22:00
DeepSeek flash official off-peak (1/4) about RMB 792 schedule into idle hours and it is half price on the spot
DeepSeek flash run straight at peak (2/8) about RMB 1,584 the baseline with no scheduling at all
Doubao Seed 2.1-Pro low-priority tier (3/15) about RMB 2,772 accept queueing in exchange for about half price
GLM-5.3 Batch 50% off (4/14) about RMB 2,904 the async tier at flagship quality

The same production line runs from RMB 227 to RMB 2,904 — a gap of roughly 13x. Put the three accounts together and the "billing-term leverage" of enterprise API buying is really just three layers: move jobs into the cheap hours the vendor officially recognizes (DeepSeek off-peak, Alibaba overnight), switch jobs from real-time to asynchronous (Batch/low-priority), and hang repeated content on the cache. All three layers are rules with prices marked in plain sight on official pages — no special channel required.

In plain terms: the order for enterprises buying API is — settle on official direct first, then let the job's "time" trade for the discount (off-peak/overnight/low-priority/Batch), and only then compete on model unit price. Pick the wrong channel layer and every saving after it is wasted.


4. Signal Interpretation

1. The price war has wound down; the competition over "how you sell" has begun. The domestic main panel has not budged for several issues, but the vendors' moves on billing methods have not stopped: DeepSeek wrote time-of-day pricing into its unit-price table, Zhipu turned Batch and limited-time free cache storage into standard equipment, Kimi offers Batch at 60% of list, and Alibaba runs an overnight discount campaign. With the open books on unit price roughly leveled, what vendors compete on is who leaves a thicker discount layer for customers who can run the numbers. For enterprise procurement this means "which vendor is cheaper" is becoming an outdated question — "which vendor is cheaper under my usage pattern" is the right one.

2. Channel price gaps are a real procurement variable. This issue's comparison shows that between official direct and aggregator platforms, the same model's input unit price can differ by as much as 2.7x and the monthly bill for the same volume of work by 2.5x (V4-Pro official off-peak about RMB 536 vs the aggregator's flat rate about RMB 1,332), while some tiers match official pricing outright (GLM-5.3). Aggregator pricing is not linked to official pricing — it may not follow official price cuts, and it may not offer the official off-peak tiers at all. One hard rule worth adding to the procurement process: for any model usage entering the formal budget, quotes are benchmarked against the official pricing page, and aggregator quotes serve only as supplementary reference.

3. Promotions and limited-time clauses cluster into expiry before year-end. The dates at hand: GPT-6.1 Sol's promo price carries an official footnote noting it runs at least through November 21; Gemini 4 Argon's promo price expires November 30 and steps up to $4/20 from December 1; Gemini 3.8 Flash's promo price expires December 31; Zhipu's GLM-5.3-Flash 50% off and its limited-time free cache storage are both limited-time policies. Before writing any "limited-time price" into a budget that spans the new year, redo the affordability math once at the post-expiry list price — when the price is low, ask one question: is this temporary.


5. Action Guide

Match your role to the checklist:

  1. Procurement/budget owners: write "official pricing page as the baseline, aggregator channels as supplement" into the model-selection review sheet; in the comparison table, leave a row each for the three columns "off-peak price/Batch price/cache price" — for any sheet where all three come back empty, budget at the full flat rate.
  2. Technical implementers: if you have already opened accounts through an aggregator and your usage is concentrated, take one week's usage and recompute the bill at official direct prices, then put the difference in front of the team — the migration cost is usually nothing more than changing one endpoint address.
  3. Batch-workload owners: move every queueable job to Batch or the low-priority tier — Zhipu at 50% off, Kimi at 60% of list, Alibaba overnight at 40% of list — trial-run one day of real tasks on each first, reconcile the bills, then switch over in full.
  4. Customer-service/Q&A owners: check your tier against Scenario 1 — if your monthly bill sits in the thousands while the conversations are plain Q&A, first check whether a flagship tier has been running lightweight work.
  5. Universal discipline: from November on, list DeepSeek's per-use web-tool fees as their own line item; never write limited-time discounts into the annual budget baseline.

Next issue (the Sunday slot, for individuals): back to the individual-perspective intelligence roundup, with topics drawn from the latest price panel.


6. This Issue's AI Tool Pick: DeepSeek's Official API Off-Peak Pricing (Not a New Model, but a "Way to Buy")

What this issue recommends is not a new model but an official way to buy that many enterprises have yet to put to use: the DeepSeek Open Platform's time-of-day pricing. deepseek-flash runs at off-peak input of RMB 1 and output of RMB 4 per 1M tokens (peak 2/8), with cache hits from as low as RMB 0.02; the flagship V4-Pro runs 4.5/13.5 off-peak. The definition of off-peak is extremely business-friendly: everything outside 9-12 and 14-18 on weekdays — evenings, the lunch break, weekends, and all day on statutory holidays.

What makes it useful: among mainstream domestic platforms, this is the tier that writes "time-slot discounts" into the formal unit-price table and has the widest off-peak coverage. Most enterprises' batch jobs (report consolidation, data processing, document review) could run at night anyway — the official channel is effectively gifting half the bill; the cache-hit price charges mere fractions for "repeated Q&A over the same knowledge base". The API is compatible with both OpenAI and Anthropic interface formats, so switching over from other models takes little rework; the concurrency cap of 2,500 (flash tier) is enough for mid-scale production jobs too.

The limitations to spell out: first, "off-peak" means jobs in the peak hours (weekday daytime) run at list price or face even tighter congestion — businesses with strong real-time requirements (daytime online customer service, live Q&A) get nothing from this discount layer, and for those scenarios the things to look at are cache and lightweight tiers; second, parameters such as 1M context and output caps differ by model tier, so check them against the official documentation before integrating; third, the official page explicitly warns that prices may be adjusted, so long-term projects should not hard-code the off-peak price as a permanent one.

Who it fits: teams whose jobs can pick their hours — batch processing, review, consolidation, production scheduling. Who it does not fit: businesses that can only respond in real time during the day, and multi-vendor setups that need several models managed from one console (that is the aggregator platform's comfort zone — but remember to reconcile the main usage back against official direct).

The bottom line: the same month's work costs RMB 1,584 run straight at peak and RMB 792 moved into off-peak — and that half-the-bill difference requires no model change and no negotiation, just one edited column in the job schedule.

⚠️ Note: all prices in this article reflect official listings as of press time; limited-time discounts, time-of-day rules, and free quotas may be adjusted — actual amounts are subject to each platform's billing.


Data note: the price data in this article was obtained from publicly available data on the respective official platforms and is the latest data at press time. Domestic prices follow the official RMB prices; international prices are quoted in USD, converted at an estimated 1 USD ≈ RMB 7.1. The ledger figures are estimates; actual costs are subject to each platform's billing.