1. Token Intelligence Brief
As we said at the close of last issue's individual-focused special, this issue returns to the enterprise and organizational perspective to run the most common bills in fine detail. As of press time, prices for the main domestic models are broadly steady: DeepSeek's two models and Zhipu's GLM-5.3 have now gone six consecutive issues without a price move, and the changes on the panel are concentrated on the "calendar" — the free window for DeepSeek's web-connected tools has only days left, and the expiry dates of several overseas promo prices are closing in. First the panel, then the accounts.
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 |
| Kimi (Moonshot AI) | Kimi-K3.1-Pro (new flagship) | 13.00 (cache hit 1.30) | 68.00 | 1M context; highspeed tier 26/136 |
| Kimi (Moonshot AI) | Kimi-K3 | 20.00 (cache hit 2.00) | 100.00 | On sale; 1M context |
| Kimi (Moonshot AI) | Kimi-K2.7-code | 6.50 (cache hit 1.30) | 27.00 | Coding-focused; closed to newly registered users from October 31 |
| Zhipu | GLM-5.3 (flagship) | 8.00 | 28.00 | Cache hit 2 RMB; storage free for a limited time; 1M context |
| Zhipu | GLM-5.3-Flash | 0.80 | 2.80 | Limited-time 50% off: input 0.4 / output 1.4 |
| Alibaba Cloud Bailian | Qwen3.8-Max | 12.00 | 36.00 | Batch at half price |
| Alibaba Cloud Bailian | Qwen3.8-Flash | 0.80 | 2.70 | New users get 1M tokens free (90 days) |
| ByteDance Volcano Ark | Doubao Seed 2.1-Pro | 6.00 | 30.00 | Cache hit 1.2; low-priority tier 3/15 |
| ByteDance Volcano Ark | Doubao Seed 2.1-Turbo | 3.00 | 15.00 | Cache hit 0.6 |
| MiniMax | MiniMax-M3.1-Flash | 1.00 | 4.00 | Cache hit 0.10 |
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. The Kimi K2 series remains on sale for existing users: the official announcement has it closed to newly registered users from October 31; already-registered users are unaffected.
B. Reference Prices on International Platforms (USD per 1M tokens)
| Platform | Model | Input | Output |
|---|---|---|---|
| OpenAI | GPT-6 Astra (flagship) | 10.00 | 50.00 |
| OpenAI | GPT-6.1 Sol (coding-focused) | 2.00 | 10.00 |
| OpenAI | GPT-6 Luna (lightweight) | 0.10 | 0.50 |
| Anthropic | Claude Opus 5.5 (flagship) | 4.00 | 20.00 |
| Anthropic | Claude Sonnet 5.5 | 2.00 | 10.00 |
| Anthropic | Claude Haiku 5.5 (lightweight) | 0.10 | 0.50 |
| Gemini 4 Argon (flagship) | 2.00 (promo) | 10.00 (promo) | |
| Gemini 3.8 Flash | 0.75 (promo) | 3.75 (promo) | |
| xAI | Grok 4.7 | 2.00 | 6.00 |
Two expiry reminders: Gemini 4 Argon's current 2/10 USD is a promotional price, officially listed through November 30, 2026, and stepping up to 4/20 from December 1; Gemini 3.8 Flash's promo price runs through December 31, 2026, after which it steps up to 1.5/7.5. If you build overseas models into long-term projects, these two dates deserve a spot on the procurement calendar.
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 |
| Anthropic | Claude Pro | About RMB 128/month billed annually, about RMB 156 paid monthly | Claude Code + more usage |
| MiniMax / Kimi | M Plan / Kimi Membership | Check after login or in the app | Independent of API pay-as-you-go billing |
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. The usual starting point for zero-cost pilots of knowledge-Q&A workloads.
- DeepSeek Search/Fetch: from November 1, 00:00, web-connected calls made inside coding tools will be billed per use (RMB 5 per 1,000 successful requests each) — it is still a free window right now, and this issue's Focus section walks through how to book that cost.
- Alibaba Cloud Bailian Night Plan: 3.8-Max/Flash at a limited-time 40% of list overnight (runs starting after 22:00) — the tool-recommendation section at the end of this issue breaks down the eligibility conditions.
- China National Supercomputing Internet: the RMB 0.1 special-offer resource pack currently contains 50M tokens (about RMB 0.002 per 1M tokens) — the theoretical low-price anchor for batch jobs; log in to the platform to verify detail billing before scaling up.
- Kimi: file-related APIs (document content extraction/storage) are free for a limited time — for long-document workloads, upload the file first and then read; the API itself carries no extra charge, and the bill only counts the model's tokens.
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: DeepSeek Web Search Billing Starts November 1 — How Should Enterprises Book the Cost
The item we flagged in passing last issue has now reached the point where action is due. DeepSeek's official pricing page currently states that Search (web search) and Fetch (web page reading) are each billed at RMB 5 per 1,000 successful requests, while the calls that reach these tools via server-side tool calls inside coding tools (DeepSeek Harness, Claude Code, Codex, etc.) are "not yet billed" — and per-call fees will start being collected from November 1, 2026, 00:00 Beijing time. In other words, the free window has only days left.
First, the nature of the change: this is not a model price increase — the flash and v4-pro tiers have gone six consecutive issues without a move; what changed is the billing method for the add-on capability of "web connectivity". It turns from an implicit freebie bundled into model calls into a separately priced, clearly tagged billing item: charged per successful request, with failures unbilled. DeepSeek is not the first vendor to give "web connectivity" its own line on the price list, but the RMB 5 per 1,000 requests pricing draws the cost scale very clearly.
The advice for enterprises is three moves, completed within the free window:
First, check existing usage. Have whoever connects the DeepSeek API through coding tools pull the past month's tool-call volume from the open-platform console — this number used to cost nothing, so chances are nobody has ever looked. It is the base for every calculation that follows.
Second, compute the increment. The formula is simple: average daily web calls ÷ 1000 × RMB 5 × 2 (Search and Fetch often fire in pairs) × 30 days. Example: a 20-person team averaging 20 web calls per person per day adds about RMB 120 a month; if automation scripts fire at high frequency — say 10,000 calls a day — that is RMB 3000 in a month, enough to overtake the entire model fee of a lightweight model.
Third, handle it by scenario. Knowledge-base workloads that already have their own retrieval built (see Scenario 2 below) find the web tools redundant — switch them off or restrict triggering, and it lands at zero cost; teams that rely heavily on web-connected coding should fold the RMB 5 per 1,000 requests into project quotes; for batch automation jobs, check whether each run truly needs the web — many "looks like it needs a search" steps drop to zero calls once replaced with a local cache or a fixed data source. Locally deployed setups and plain API chat users are unaffected.
The bottom line: November 1 is a dividing line in cost structure, not a price-hike notice. Whoever maps their own call volume first is far less likely to be startled by this new bill.
3. Token Ledger: The Three Accounts Enterprises Run Most Often
If you have ever wondered "now that the company is on AI, what does the API actually cost per month" — find your account below. Everything here is an estimate: scenarios are built from typical business volumes, and actual costs float with the usage mix; international prices are converted at 1 USD ≈ RMB 7.1. The three accounts map to the three most common consumption patterns in enterprises: "reads a lot", "asks often", and "runs late".
Scenario 1: Bid & Contract Review (100 Long Documents a Month) — Is "1M Context" Worth It
Setup: each document is about 400K tokens (a mixed Chinese-English bid of roughly 600K characters), output is a key-point extract of about 3000 tokens, 100 documents a month.
| Plan | Estimated monthly cost | Reading approach |
|---|---|---|
| GLM-5.3-Flash 50% off (0.4/1.4) | about RMB 16 | segmented read |
| DeepSeek flash off-peak (1/4) | about RMB 41 | segmented read |
| DeepSeek flash peak (2/8) | about RMB 82 | segmented read |
| DeepSeek v4-pro off-peak (4.5/13.5) | about RMB 184 | segmented read |
| GLM-5.3 (8/28) | about RMB 328 | 1M whole-read |
| Kimi K3.1 Pro (13/68) | about RMB 540 | 1M whole-read |
| Gemini 4 Argon promo price ($2/$10) | about RMB 589 | 1M whole-read |
| Kimi K3 (20/100) | about RMB 830 | 1M whole-read |
For the same batch of documents, the bill runs from RMB 16 to RMB 830 — a gap of roughly 51x. The source of the gap is not "which vendor is cheaper" but the reading approach: for a contract running 400K tokens a copy, reading it in two passes with a 200K-context lightweight model, even with a 10% segment-overlap overhead added in, scales the Flash 50%-off basis only up to about RMB 18; what "1M whole-read" buys is "never cut apart" — clause conflicts across chapters and cross-document reconciliations are visible only when the whole volume is read as one. Most bid reviews aim to extract key elements, and segmented reading is enough; step up to the 1M tier only when whole-volume reasoning is genuinely needed (for instance, checking whether earlier and later clauses inside one bid contradict each other). One more note: on the same 1M whole-read, K3.1 Pro saves a third over K3 — the "new flagship price cut" we covered last issue shows up even more clearly in an enterprise's batch bill.
Scenario 2: Enterprise Knowledge Base Assistant (50-Person Team) — Monthly Bill of RMB 18 or 16,000
Setup: a 50-person team asks a combined 6600 questions a month, with answers of about 800 tokens. Two technical routes: the RAG route uses retrieval to pull the relevant passages first and carries only 4000 tokens into the model per question; the long-context route stuffs the full text — 300K tokens of material — into the model every time. The model-fee comparison:
| Route | Plan | Estimated monthly cost |
|---|---|---|
| RAG | GLM-5.3-Flash 50% off (0.4/1.4) | about RMB 18 |
| RAG | GLM-5.3-Flash standard price (0.8/2.8) | about RMB 36 |
| RAG | DeepSeek flash off-peak (1/4) | about RMB 48 |
| Long context | GLM-5.3 (8/28) | about RMB 15,988 |
| Long context | Kimi K3.1 Pro (13/68) | about RMB 26,099 |
| Long context | Gemini 4 Argon ($2/$10) | about RMB 28,491 |
Same company, same questions, and the two architectures differ by roughly 1587x. That is not hyperbole; it is the math of input structure: RAG carries 99% less content per question, so the model fee falls off a cliff. The long-context route is not unchoosable — for a small team with single-digit daily questions, documents that keep changing, and no desire to maintain a retrieval library, spending a few thousand RMB for the peace of mind of "never retrieving wrong" is a coherent trade. But for daily Q&A at 50-person scale, the one-off engineering investment in retrieval buys a three-orders-of-magnitude gap in the monthly bill. When choosing a knowledge-base architecture, ask about the setup first and the unit price second.
Scenario 3: Night Batch Production Line (264M Input / 132M Output a Month) — The Two-Layer Discount of Time Slots and Batch
Setup: a content-production/data-processing pipeline, with monthly input of about 264M and output of about 132M tokens.
| Plan | Estimated monthly cost | Notes |
|---|---|---|
| GLM-5.3 Batch 50% off (4/14) | about RMB 2,904 | flagship quality + async half price |
| DeepSeek flash run in daytime peak (2/8) | about RMB 1,584 | no-scheduling baseline |
| OpenAI GPT-6 Luna Batch 50% off ($0.05/$0.25) | about RMB 328 | overseas lightweight benchmark |
| DeepSeek flash off-peak scheduling (1/4) | about RMB 792 | move to idle hours and the bill halves |
| Alibaba Qwen3.8-Flash overnight 40% of list (0.32/1.08) | about RMB 227 | Night Plan terms, see Section 6 |
The do-nothing baseline is RMB 1,584 — same work, nothing swapped out, just move the jobs out of working hours (DeepSeek's idle hours = all times outside the two weekday peaks, 9-12 and 14-18, including midday, evenings, and all day on weekends and holidays) and the bill halves on the spot to RMB 792; stack a lightweight model or half-price Batch on top and it steps down further. The range runs from RMB 2,904 (flagship Batch) as the quality ceiling, to RMB 792 (off-peak scheduling) as the easy scheduling win, to RMB 227 (lightweight + overnight 40% of list) as maximum compression — quality requirements set the ceiling; scheduling habits set the floor. As for the theoretically cheaper resource-pack basis (the supercomputing special offer works out to about RMB 0.002 per 1M tokens — the same volume would theoretically cost one or two RMB), detail billing must be verified after logging in to the platform; treat it here only as a price anchor for reference, not as a figure to budget against directly.
In plain terms: the order for enterprises to save on API costs is — fix the input structure first (RAG/segmentation), then fix the time slots (off-peak/night), and only then change models. Run all three steps and most scenarios' bills drop from the "tens of thousands" tier to the "hundreds" tier.
4. Signal Interpretation
1. Tool calls are becoming a billed item; web connectivity is no longer the model's free gift. By pricing Search/Fetch separately and charging per successful request, DeepSeek is splitting "AI + web" into two SKUs. For enterprise budgets the implication is: from now on the quote needs a line item for "tool call fees". Other vendors are expected to follow with their own itemized terms — when comparing procurement options, put tool pricing into the comparison table alongside model unit prices.
2. The source of bill gaps is shifting from "unit price" to "usage". The most eye-catching numbers in this issue's three accounts have nothing to do with price hikes: Scenario 1's 51x and Scenario 2's 1587x are both determined by reading approach and architecture. Over the past year model unit prices have declined across the board (flagship tiers fell from over a hundred RMB per 1M tokens to a few dozen, and lightweight tiers entered the 1-RMB range); the dividend on the unit-price side has been diluted, and the main battlefield of enterprise cost optimization from here is input structure and scheduling — which is why "knowing how to run the numbers" is becoming the core competitive skill of AI projects.
3. Calendar-driven cost events are clustering; schedule procurement against the dates. Near-term expiries: October 31, the Kimi K2 series closes to newly registered users (the migration window for existing users); November 1, DeepSeek's web-connected tools start billing; November 30, the Gemini 4 Argon promo price expires and doubles; December 31, the Gemini 3.8 Flash promo price expires and doubles. A promo price's "low" is a limited-time "low" — if you build overseas promo models into long-term projects, redo the renewal math at the post-expiry prices.
5. Action Guide
Match your role to the checklist:
- Do it this week (every team on DeepSeek): pull the past month's web-call volume and estimate the increment from November using "calls ÷ 1000 × RMB 5 × 2 × 30 days"; for scenarios that do not need the web, switch off triggering before November 1, and for high-frequency scenarios that do, write the cost into the November budget.
- Long-document workloads: default to segmented reading + a lightweight model to start (the RMB 16-82/month tier), and step up to 1M whole-read (the RMB 328-830/month tier) only after confirming cross-document/whole-volume reasoning needs; before signing, run a sample through both reading approaches and set the tier by measured results, not by name value.
- Knowledge-base projects: start new projects on the RAG route by default; if you are already on direct long-context with high daily question volume, take this issue's Scenario 2 comparison table into a meeting with the IT team — that bill gap is enough to fund a solid project bonus.
- Batch production lines: queue every queueable job into off-peak/night hours, and route through Batch wherever Batch works; trial-run with one week's volume before scheduling production, and never budget straight off the theoretical lowest price.
- Universal discipline: put the four dates — October 31, November 1, November 30, December 31 — on the team calendar, and re-verify prices ahead of each expiry.
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: Alibaba Cloud Bailian Night Plan (Overnight 40% of List)
After running the three accounts, this issue's pick is an official program that "puts a discount on the bill" — Alibaba Cloud Bailian's Night Plan overnight offer: calls to qwen3.8-max and qwen3.8-flash that start after 22:00 run at 40% of list (3.8-Flash works out to about RMB 0.32/1.08 per 1M tokens after the discount) — among the strongest time-slot discounts currently on offer through major vendors' official channels.
Why it is worth a look: it is an official platform program, not a reseller's promo price, and the billing terms sit right on the official billing page; the 40% rate is a real bill reduction for overnight batch jobs, and it does not conflict with "switching to a smaller model" — for teams that want to keep quality while saving money, running max overnight at 40% of list may land closer to the desired result than running flash in the daytime.
The limitations and pitfalls that must be stated: first, the overnight tier has a time threshold — it applies only to runs starting after 22:00, and jobs hung up in the daytime run at full price for nothing; second, the campaign page lists the applicable scope as specific client scenarios such as Qoder and Meoo, and whether generic API calls qualify, and which models are covered, is subject to the campaign page's real-time listings — before scheduling production on it purely for the 40% rate, verify the bill with a small batch of calls first; this is the pitfall this issue most wants to flag; third, limited-time programs can change at any moment, so do not write the 40% rate into the annual budget baseline for long-term projects.
Who it fits: teams that can move work to the night — content production, data processing, daily-report consolidation, batch code review. Who it does not fit: businesses that must respond in real time during the day (customer service, live Q&A) — the overnight discount has nothing to do with you, and effort spent on RAG and caching pays back more.
The bottom line: the same job run straight in the daytime costs RMB 1,584; scheduled into the late night, RMB 792; overnight at 40% on the lightweight tier, RMB 227 — the leverage for saving money has never been on the unit-price table; it lies in what hour you schedule the job and which tier you run it on.
⚠️ Note: the applicable scope, model list, and plan terms of the overnight discount are all subject to the real-time listings on each official page; promotions and discounts may change.
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.
广告


