Coding Plan: Subscriptions and Billing
📚 Series Navigation: The previous article 05 Accessing Third-Party / Domestic Models taught you how to connect DeepSeek, Kimi, GLM into Claude Code. This article calculates the bills clearly—which plan to subscribe to, how the money is spent, how to check usage, how to save.
There is a number in the official documentation that might make you pause at first glance: in enterprise deployments, the average developer burns through roughly $13 per active day, or $150–250 per month (Source: official costs documentation). $13 a day, by exchange rate is over ninety RMB, almost two thousand a month—and this is just "average."
I myself suffered a loss when I first used Claude Code connected to the official API (pay-as-you-go): without installing any usage prompts, I just buried my head and worked for a week. Opening the Console on the weekend, I looked at the bill and it was over $40—the money wasn't a lot, but I couldn't explain how it was spent at all: which conversation burned the most? Did I choose the wrong model or pack the context too full? A messy account.
This article comes to calculate the account clearly. Understanding the billing logic is much more important than blindly finding a cheap plan—the same plan, someone who knows how to use it lasts a month, someone who doesn't hits the ceiling in three days.
After reading this article, you will get:
- A "Pay-as-you-go vs Subscription Plan vs Domestic Coding Plan" comparison table, to know which path you should choose
- Understand Claude Code's billing unit (token), to know exactly where the money is spent
- A few commands (
/usageto check usage,/usage-creditsto set limits,/statusto verify config), so you can keep an eye on expenses yourself - A set of money-saving habits tested effectively, for the same work, tokens burned can be cut by a large chunk
01 First Understand: What Are You Paying For
Conclusion first: Claude Code itself is free, you pay for the "model compute".
Claude Code is just a command-line tool running in the terminal, downloading is free. What really costs money is the large model behind it—when you ask a question, it reads a file, it writes code, all content must be sent to the model for processing, this processing volume is the billing unit.
Analogy: Claude Code is a taxi, the model is the mileage. Getting in the car is free, the car is free for you to sit in, but once it starts driving every kilometer it runs is on the meter—the more you ask, the larger the files you let it read, the longer it thinks, the longer the "mileage".
The unit of measurement for this "mileage" is called token. Analogy: token is the "pricing word count" for text. The model cuts text into small chunks to count—according to Anthropic's official platform glossary, for Claude, 1 token represents approximately 3.5 English characters, exact numbers vary by language (no fixed ratio given for Chinese); what you send out counts as "input tokens", what comes back counts as "output tokens", priced separately, output is usually more expensive.
To give a few scenarios you encounter every day, feel how tokens pile up:
- Tell it to "read this 2000-line file before modifying" — the whole file becomes input tokens sent out
- A conversation round reaches the 50th message — the previous 49 messages are resent every time as context (this is the big chunk, the
/clearhabit in section 05 is aimed at this) - Enabled extended thinking, it "thinks" a large paragraph before acting — the thinking process counts as output tokens
So how much money is spent has little to do with "how many questions asked," but has immensely to do with "how much context is stuffed each time"—this is the core grip for saving money later.
💡 One-sentence summary: The tool is free, you pay for the model's compute fee billed by tokens—input + output + thinking, all cost money.
02 How to Choose Among Three Paths: Pay-As-You-Go / Official Subscription / Domestic Coding Plan
Knowing that you pay for tokens, the next question is: through what channel is this money paid? There are three mainstream paths, let's show the table first then speak plainly.
| Plan | How it's Charged | Monthly Cost (Ref) | Suitable For | Main Pitfalls |
|---|---|---|---|---|
| Official Pay-As-You-Go (API / Console) | Billed by token usage, deduct as you use | Floating, heavy users $150–250+ | Want to use official native model, stable usage | Bill floats, might exceed budget; needs card bind (USD) |
| Official Subscription (Pro / Max / Team) | Fixed monthly fee, includes certain quota | Fixed (subject to official pricing) | High frequency use, want fixed expense, use official model | Overseas payment, needs proxy; quota has cap |
| Domestic Coding Plan (Ark / Bailian / iFlytek etc.) | Fixed monthly fee, quota shared by multiple tools | Low, first month often has promos | Domestic users, sensitive to cost, can accept domestic models | Model not Claude; quota per vendor rules; prices volatile |
These three paths are not about "which is best," but "who you are." Two points added outside the table:
The two official paths—pay-as-you-go uses native Claude models (first tier for programming), most flexible but bill floats (that messy $40 earlier was this); subscription has the least psychological burden, Max suits heavy users, but requires overseas payment + proxy, how much for the tiers always subject to official claude.com/pricing, won't be hardcoded here.
Domestic Coding Plan is the most practical choice for domestic developers right now. Volcengine Ark, Alibaba Bailian, iFlytek Spark all launched "Coding Plans"—subscription services specifically designed for AI programming tools, fixed monthly fee, far lower than pay-as-you-go, compatible with Claude Code, integrating Qwen, Kimi, GLM, DeepSeek, Doubao, MiniMax, multiple tools (Cursor, Cline, etc.) sharing one quota.
⚠️ Domestic Coding Plan price promotions are extremely frequent, rumors online like "first month 9.9 / 49.9" are campaign prices for a certain period, will change at any time, campaigns will be taken down. This article does not quote specific amounts, please go directly to the vendor's campaign page to see current prices.
The access method (Base URL, ANTHROPIC_MODEL) has been thoroughly explained in the previous article, won't repeat. But there is one major billing pitfall that must be singled out, specifically talked about in the next section.
💡 One-sentence summary: Official native choose subscription or pay-as-you-go, domestic saving money choose Coding Plan—first think clearly who you are, then pick the path, don't just look at the price.
03 The Billing Pitfall of Domestic Coding Plans: Wrong Base URL, Money Wasted
Singled out separately, because too many people fall here—clearly bought a Coding Plan, but receive an extra bill at the end of the month. The root cause is: the same vendor often has two sets of Base URLs, one uses plan quota, one uses pay-as-you-go, and they look very similar. Take Volcengine Ark as an example:
| Base URL | Which Money it Takes | Purpose |
|---|---|---|
https://ark.cn-beijing.volces.com/api/coding | ✅ Consumes Coding Plan quota | Connecting Claude Code should use this |
https://ark.cn-beijing.volces.com/api/v3 | ❌ Does not use plan, bills pay-as-you-go separately | Regular API calls |
Do you see it? The difference is just the one word between coding and v3. You bought a plan, thought the fixed monthly fee capped it, but your hand slipped during configuration and filled in /api/v3, the plan quota didn't move at all, all calls went to pay-as-you-go, paying extra out of pocket at the end of the month. Alibaba Bailian is similar—its Coding Plan exclusive Key (starts with sk-sp-) and pay-as-you-go regular Key (starts with sk-) are not interoperable, mixing them also causes problems.
So connecting a domestic Coding Plan, remember two iron rules: Base URL must contain the word coding; API Key use the one exclusive to the plan, don't put the pay-as-you-go Key on top. The specific addresses and Key prefixes for Ark and Bailian above are private rules of third-party vendors, will change as vendors adjust, please refer to the current instructions of each vendor's official documentation / console before acting.
If you mistakenly fill in v3 when configuring Volcengine Ark, if you habitually run /usage that night to check usage, you'd find "why didn't the plan quota move, but the Console pay-as-you-go side has numbers," and quickly change back to coding. This kind of pitfall is hard to discover by eyeballing the configuration, can only be caught by watching usage.
💡 One-sentence summary: Bought a plan ≠ using the plan—Base URL choose one with
coding, Key use plan exclusive, verify immediately after configuring, otherwise money quietly flows away from another opening.
04 How to Check Usage: Three Commands to Watch Your Wallet
Paying money without looking at usage is equal to driving blindfolded. Claude Code has built-in a few commands letting you know where the money is spent at any time, discussed one by one.
/usage: The Most Complete Usage Panel (Used Most Frequently)
Type directly in the dialogue box:
/usageIts top Session block displays detailed token usage of the current session, looks like this (Source: official costs documentation):
Total cost: $0.55
Total duration (API): 6m 19.7s
Total duration (wall): 6h 33m 10.2s
Total code changes: 0 lines added, 0 lines removedThe four lines are: estimated total cost, actual API processing duration, true duration the window was hanging, lines of code changed. After finishing a task just knock it casually, to have a number for the current expense. But the official has an important reminder:
The dollar figure in the Session block is estimated locally from token counts, and may differ from your actual bill. Authoritative billing looks at the usage page on Claude Console. (Source: official costs documentation)
If you use Official Pro / Max / Team / Enterprise Subscriptions, /usage will have an extra block of quota usage breakdown—breaking down recent consumption by skills, subagents, plugins, each MCP server showing proportions, press d / w to toggle "past 24 hours / 7 days". Note this is an approximation calculated from local session history, does not include usage from other devices or claude.ai. (Source: official costs documentation)
/usage-credits: Set a Monthly Limit for Yourself (Anti-Overspend Artifact)
This is a feature exclusive to official Pro / Max plans (Source: official costs documentation):
/usage-creditsIt allows you to set a monthly spending limit on usage quota. Once you spend up to this number, Claude Code will prompt you "Raise limits or remove limits", decide directly in the CLI. Changing this limit requires billing access to the account.
Beginners are strongly recommended to use it right away to set a limit—for example, a budget of 100 bucks a month, it blocks you when you hit the roof, much better than getting a scare looking at the bill at the end of the month. My messy $40 was precisely because I completely didn't know this command existed at the time.
There is also a /status, opening the Status tab in the settings interface, according to the official commands documentation displays version, model, account, connectivity. The official recommends in the Third-Party Integrations documentation to "use /status to verify your proxy and gateway configuration has been applied correctly"—it is precisely the tool to check access configurations from the previous section, after configuring the plan knock it, confirm model and connectivity are normal before working.
| Command | Applicable Scenario | Notes |
|---|---|---|
/usage | Check current session token usage and estimated cost | Amount is local estimate, authoritative bill look at Console |
/usage-credits | Set monthly usage limit, prevent overspend | Only available for official Pro / Max plans |
/status | Verify version, model, account, connectivity | The first step to verify if access configuration took effect |
💡 One-sentence summary:
/usageto see breakdown,/usage-creditsto set limit,/statusto verify config—three commands weld your wallet shut.
05 The Real Skill in Saving Money: Let the Same Work Burn Less Tokens
Changing to a cheap plan is the bad policy, knowing how to use the tool is the good policy. Back to that sentence in section 01—how much money you spend is extremely related to how much context you stuff in, so the core of saving money is just one sentence: Don't let the model read useless things, don't let it think unnecessary detours. The following are all from the official costs documentation, picking the ones most immediately effective for beginners.
First, clear between tasks with /clear. This is the habit with the highest return.
/clearI myself used to open a window from morning till night, fixing bugs in the morning and writing new features in the afternoon all squeezed into one session, too lazy to switch. The problem is context accumulation: in the afternoon, every question resent that bunch of dialogues from the morning which were long useless, tokens doubled for nothing—later staring at /usage I found the unit price of those few dialogues in the afternoon was obviously higher than in the morning, that's how it burned. Official original words:
Use
/clearto start fresh when switching to unrelated work. Stale context wastes tokens on every subsequent message.
Develop the habit: immediately /clear after completing an independent task, never hard-wire to old sessions when switching to unrelated work. Just this one move, the expense of long sessions is cut by a large chunk.
Second, choose the right model, don't use a cannon to hit a mosquito. Official principle is "Sonnet handles most programming tasks excellently and costs less than Opus, reserve Opus for complex architecture and multi-step reasoning" (Source: costs documentation). Speaking plainly: daily modifying code, writing functions, debugging small bugs, Sonnet is totally enough; only for hard bones like "refactoring an entire module" use /model to cut to Opus. Simpler subagent tasks can even specify Haiku in the subagent configuration, lower unit price. Default hanging Sonnet, a week only cut two or three times.
Third, use plan mode first for complex tasks. Press Shift+Tab to enter, let it explore the codebase first, propose a plan for you to approve before acting—rework due to wrong direction burns the most money. This pitfall is very common: let it modify a poorly explained requirement, it buries its head modifying seven or eight files, direction totally wrong, this stretch of tokens goes down the drain.
Fourth, write prompts specifically. The vaguer you are, the more it has to flip through files all over the world guessing what you want to do, every flip is tokens. Official original words comparison: "Improve this codebase" triggers a broad scan, "add input validation to the login function in auth.ts" lets it work efficiently with minimal file reads.
The table below can be treated as a "Money Saving Quick Reference" stuck next to your monitor:
| Bad Habit ❌ | Good Habit ✅ | Saves Where |
|---|---|---|
| One session from morning to night | /clear between tasks | Doesn't resend stale context |
| Hang on Opus the whole time | Default Sonnet, hard problems then Opus | Lower unit price |
| Vague requirement act directly | Complex tasks plan mode first | Less rework, less blind scanning |
| "Help me improve the code" | "Add input validation to login func in auth.ts" | Read less irrelevant files |
Trivia (Source: official costs documentation): Claude Code will also use a tiny bit of tokens when idle—for example, summarizing old conversations in the background for claude --resume. But the amount is very small, typically less than $0.04 per session, no need to worry.
💡 One-sentence summary: Saving money is not changing to cheap plans, it is clear frequently with
/clear, Sonnet as baseline, plan first for complex, state prompts clearly—the smaller the context, the thinner the bill.
06 Hands-on: Three Steps to Watch Your Expenses This Week
Reading without practicing is equal to not reading. This set of actions below you can run through right now, establishing "usage awareness" in five minutes.
Prerequisite: Already installed and configured Claude Code according to previous articles (official subscription or domestic Coding Plan both fine). Launch in project directory:
cd path/to/your_project
claudeStep One: Verify configuration didn't fill wrong opening.
/statusIt will open the Status tab of the settings interface, according to official commands docs displays version, model, account, connectivity—key is to confirm model is the one you want to use, connectivity normal, otherwise it means the access wasn't configured right.
Base URL itself is not in the output of /status, you have to go back to check the environment variable you wrote: domestic Coding Plan users confirm ANTHROPIC_BASE_URL contains the word coding (e.g. .../api/coding), not a pay-as-you-go address like /api/v3. If filled wrong go back to that set of env vars in Article 05 to correct it, reopen terminal and re-launch.
Step Two: Randomly ask a question, see how much was spent. First let it do a small job:
Help me see what files are in the current directory, briefly say what this project doesAfter it finishes replying, type:
/usageThe top Session block is expected to look like this:
Total cost: $0.0X
Total duration (API): Xs
Total duration (wall): Xs
Total code changes: 0 lines added, 0 lines removedRemember this number. This is the cost of "asking one simple question + reading the directory once"—with a baseline, you'll be able to spot immediately if a future conversation is abnormally burning money.
Step Three: Set monthly limit (Official Pro / Max Users).
/usage-creditsFollow the prompts to set a limit you can accept. After setting, it will actively block you when hitting the roof, asking if you want to raise it, equal to putting a safety catch on your wallet.
Domestic Coding Plan users:
/usage-creditsis targeted at official usage quota, may not apply to your vendor plan—quota and usage subject to vendor console, but habits like/usage,/clearshould still be developed exactly the same.
After running these three steps, you upgrade from "blindly burning money" to "knowing exactly how much burned at any time, and can set a gate".
💡 One-sentence summary:
/statusverify opening →/usageestablish baseline →/usage-creditsset limit, take back control of expenses in five minutes.
07 Summary
This article broke down the "money" of Claude Code and explained it thoroughly.
| What You Learned | Key Points |
|---|---|
| Why Pay | Tool free, paying for model's compute fee billed by tokens |
| How to Choose 3 Paths | Official native→subscription/pay-as-you-go; Domestic save money→Coding Plan |
| Pitfalls of Domestic Plans | Base URL choose with coding, Key use exclusive |
| How to Check Usage | /usage see breakdown, /usage-credits set limit, /status verify config |
| How to Save Money | /clear clear freq, Sonnet baseline, complex plan first, state prompts clearly |
If you only remember one thing: How much money is spent is extremely related to "how much context stuffed in", not to "how many questions asked". Digest this sentence, and you can deduce all the money-saving tricks yourself—understand the billing logic, check current expenses, set monthly limits, do work more economically, by here you have mastered it all.
⚠️ Emphasizing accuracy once again: The plan prices, promotions, quota rules mentioned in this article will change, amounts are always subject to official claude.com/pricing or domestic vendor campaign pages; command behaviors like /usage-credits subject to official costs documentation, may update with versions (claude --version to see version).
The accounts are clear, the plan is chosen, next article 07 "First Use: Running Through the First Example"—officially launching Claude Code, starting from a minimal real project, running through the "Think → Act → Observe" agentic loop with your own hands. Enough buildup, it's time to get real.