Should one developer use a Claude subscription or an API key for Claude Code?

Claude Code is Anthropic's coding tool. You type a request in a terminal, the text window where you type commands, and an AI model reads files, runs commands and edits code for you. There are two ways for one developer to pay for it. The first is a Claude subscription, a Pro or Max plan, which includes Claude Code and counts the work against a usage limit: an allowance of work the plan includes, which Anthropic states as a multiple of another plan's allowance, with a session window that resets every five hours and a weekly window. The second is an API key, the credential of a Claude Console account, which is Anthropic's account site for developers who pay for the Claude API per token, a token being one of the pieces of text the model reads and writes. This page is for the developer choosing between them, and it is part of the guide to Claude Code usage limits. Every fact comes from Anthropic's documentation and support articles read on 4 October 2026, and Anthropic's figures are its own statements about its own product.

What a subscription includes, and when it stops

Anthropic's support article on using Claude Code with a Pro or Max plan says that with those plans "you now have access to both Claude on the web, desktop, and mobile apps and Claude Code in your terminal with one unified subscription" [4]. The pricing page, read on 4 October 2026, lists Pro at $17 per month with an annual subscription of $200 billed up front, or $20 per month billed monthly, and Max from $100 per month, with a choice of five times or 20 times the usage of Pro; Anthropic says prices may change [6]. The page on what each plan includes sets the plans side by side.

The subscription stops when a window runs out. The support article says both plans "offer usage limits that are shared across Claude and Claude Code", that you see warning messages about remaining capacity, and that when you reach the limit you can upgrade, enable usage credits, switch to a Claude Console account and buy API usage, or wait for the reset [4]. The page on the five-hour window and the weekly window covers the two windows, and the page on waiting for a reset covers the wait.

One more sentence matters for this choice. Anthropic's Pro plan article says: "The Pro plan does not include API usage through the Claude Console. If you're interested in both enhanced Claude features and the Claude API, you'll need to set up Console access to pay for API usage separately" [7].

What an API key does

Anthropic's cost documentation opens with the rule: "Claude Code charges by API token consumption" [1]. With an API key, you pay for every token at the list prices on Anthropic's pricing page, read on 4 October 2026, in US dollars per million tokens [2]:

Model	Input	Five-minute cache write	Cache read	Output
Claude Opus 5.5	$4	$5	$0.20	$20
Claude Sonnet 5.5	$2	$2.50	$0.20	$10
Claude Haiku 4.5	$1	$1.25	$0.10	$5

A cache read is a token the prompt cache served without processing it again. The prompt cache is a store, kept by Anthropic's service, of request text it has already processed. Anthropic's pricing page says a cache read costs 0.1 times the base input price on most models and 0.05 times on Opus 5.5 [2]. In the same table, a one-hour cache write is 2 times the base input price and a five-minute write 1.25 times; those two ratios are our arithmetic on the table's figures.

Instead of plan windows, an API key has rate limits: upper limits on how many requests and tokens you can send in a period. Anthropic's pricing page says they "vary by usage tier", and names three tiers: Start, "entry-level limits for getting started"; Build, "increased limits for growing applications"; and Scale, "highest standard limits for production workloads". The numbers are on a separate rate limits page that is outside this guide's sources [2]. The support article describes the plan and the Console as "two distinct systems" [4].

The cache lifetime differs too. Anthropic's prompt caching documentation says Claude Code requests the one-hour cache lifetime for the main conversation only on a Claude subscription within the plan's included usage, and that on usage credits, an API key or a cloud provider the default is five minutes [3]. On an API key you can set promptCacheTtl to 1h to give the main conversation a one-hour cache; the setting requires Claude Code v2.1.242 or later, and the one-hour write is billed at the higher rate above [3]. The page on the cache lifetime of five minutes or one hour explains when that is worth paying.

Two more facts from the cost documentation. Background work costs tokens on either sign-in, and Anthropic puts it at typically under $0.04 per session [1]. And when you first authenticate Claude Code with a Console account, Anthropic creates a workspace named Claude Code in that account for cost tracking, for which you cannot create API keys [1].

How the two sign-ins interact

The support article warns about one setting: if an ANTHROPIC_API_KEY environment variable is set on your system, Claude Code uses that key instead of your subscription, and the result, in the article's words, is API usage charges in place of the subscription's included usage [4]. An environment variable is a named value your operating system makes available to programs. If you are logged in through a Console pay-as-you-go account and want the subscription, run /login inside Claude Code to switch [4].

Claude Code can also offer API credits when a plan limit is reached. The article says usage through that option is billed at standard API rates, distinct from Pro and Max plan pricing, that auto-reload is a Console setting, and that "all transitions to API credit usage require explicit user consent" [4]. To stay inside the plan, decline the option and let the usage period reset; to stop the option appearing at all, run claude logout, then claude login with only your plan credentials and no Console credentials [4]. Usage credits are a separate feature on the subscription itself, billed at standard API rates and charged separately from the subscription [5]; the page on usage credits after the limit explains them.

Anthropic's own average, as a point of comparison

Anthropic's cost documentation gives one set of figures, read on 4 October 2026: across enterprise deployments, Anthropic states an average of $13 per developer per active day and $150 to $250 per developer per month, with costs below $30 per active day for 90 percent of users [1]. Three things to remember. The figures are Anthropic's own, about enterprise deployments, and the same paragraph says costs "vary widely based on model selection, codebase size, and usage patterns such as running multiple instances or automation" [1]. They say nothing about a subscription's allowance, because Anthropic gives no conversion between an allowance and tokens or dollars. And Anthropic's advice is to measure before deciding: start with a small pilot and use the tracking tools to establish a baseline [1]. The guide to Claude Code costs for teams treats the same figure from a company's side.

How to decide from your own /usage

The deciding number is yours, and Claude Code already shows it. Anthropic's cost documentation says the Session block at the top of /usage "shows API token usage and is intended for API users", that subscribers have usage included so "the session cost figure isn't relevant for billing purposes", and that Claude Code "computes the dollar figure locally from token counts at list price" [1]. Reveneau reads those three sentences together: on a subscription, the Session block's dollar figure is what the same tokens would have cost at list price on an API key, and that is the comparison a developer needs. The documentation's example line is claude-sonnet-4-6: 1.2k input, 5.3k output, 940.0k cache read, 50.0k cache write ($0.55) [1]. The totals reset when /clear starts a new session, on v2.1.211 or later, so add the figure from each session across a day [1]. The figure is an estimate, and Anthropic says the Console's Usage page is the authoritative record of billing [1].

The method, then: for a month, note the Session block figure at the end of each session on your subscription, add them up by day, and compare the monthly total with what the plan costs you. If the total at list price is lower than the plan price month after month, the API key is cheaper for your pattern of work. If it is higher, the plan is cheaper, and you accept the plan's windows in return. The page on how to read /usage and /context covers each row of the screen.

An invented example at list prices

The following example is invented. Its token counts come from no source, and it exists only to show the arithmetic with Anthropic's list prices read on 4 October 2026 [2]. Suppose one active day of Claude Code work sends 200,000 uncached input tokens, writes 200,000 tokens to the five-minute cache, reads 1,800,000 tokens from the cache, and produces 100,000 output tokens.

On Sonnet 5.5: 200,000 × $2 per million is $0.40; 200,000 × $2.50 per million is $0.50; 1,800,000 × $0.20 per million is $0.36; 100,000 × $10 per million is $1.00. The day costs $2.26, and 20 such days cost $45.20.

On Opus 5.5: 200,000 × $4 per million is $0.80; 200,000 × $5 per million is $1.00; 1,800,000 × $0.20 per million is $0.36; 100,000 × $20 per million is $2.00. The day costs $4.16, and 20 such days cost $83.20.

Against the pricing page's Pro price of $20 per month billed monthly [6], the invented Sonnet 5.5 pattern costs more on an API key and the plan is the cheaper choice, if the plan's windows allow that much work; whether they do is a fact about your account that no source states in tokens. Change the invented counts to your own and the answer can reverse. That is the point of the example: the arithmetic is simple, and only the token counts are difficult to know, which is why the previous section tells you where to read them.

Subscription or API key, by what matters to one developer

What matters to you	Pro or Max subscription	Console API key
How you pay	A fixed monthly price; Pro $20 monthly or $17 on an annual plan, Max from $100 [6]	Per token at list prices, billed on actual monthly usage [1][2]
What stops you	The five-hour session window and the weekly window, shared across Claude and Claude Code [4]	Rate limits that vary by the Start, Build and Scale usage tiers [2]
What happens at the stop	Wait for the reset, use usage credits, upgrade, or switch to a Console account [4]	Reduce your request rate or request a higher tier; usage is metered per request [1][2]
Prompt cache lifetime on the main conversation	One hour within plan usage; five minutes on usage credits [3]	Five minutes by default; promptCacheTtl set to 1h on v2.1.242 or later buys the hour at the higher write rate [3]
Where to read spend	The plan usage bars in /usage and Settings > Usage on claude.ai [1][4]	The Console Usage page for the bill, and /usage for the estimate [1]
Which /usage block matters	The plan usage bars; the Session block shows what the work would cost at list price [1]	The Session block [1]
Access to the Claude apps	Included in the same subscription [4]	A Console account covers the API only; the Pro plan does not include Console usage [7]
Effect of ANTHROPIC_API_KEY in your environment	Claude Code uses the key instead of the subscription and charges API usage [4]	This is the normal sign-in

Which to pick

Reveneau is an AI software development consultancy, and all of its code is written by AI, so this choice is one every Reveneau build makes, and Reveneau is independent of Anthropic. Its recommendation from the facts above is to start on a subscription, because the price is fixed, the Claude apps come with it, and the one-hour cache lifetime applies without any setting. Read the Session block in /usage for a month. Move to an API key, or add a Console account beside the plan for the periods of intensive work that Anthropic's own article names [4], only when that month shows the list-price total below the plan price, or when the plan's windows interrupt you more often than your work can take. Anthropic's $13 per active day average is a figure about other people's deployments and belongs in the decision only as a reference [1]. The rules on this page are as Anthropic published them on 4 October 2026, and the page on the dated log of limit changes records what has changed since.