Claude Code costs more the more you use it: what an AI coding agent costs per month, and how to bring the bill down

Inside the allowance, Claude Code is a flat monthly fee. Pro is $20, or $200 billed once a year, which works out to $17 a month. Max is monthly only: $100 for 5x and $200 for 20x. After the five-hour window or the weekly limit, usage credits bill every further turn at API list price. The list price does not discount for the plan. Moving the same turn from Haiku to Fable can multiply it by about 100.

Today is 19 October 2026. Plan prices are from claude.com/pricing opened that day. Per-million-token rates are from the rate card PDF effective 7 October, and they match the model cards on the pricing page. How the allowance is shared with chat, and how overage is billed, is from Claude Code with Pro or Max and usage credits. The June note that brought one bill from $800 back to $150 is still useful. Its token rates belong to June. Do not multiply this month’s turns by them.

The bill has two parts

The first part is the seat. Pro or Max covers claude.ai on the web, desktop, and mobile, and Claude Code, from one allowance. The terminal and the Claude Code extension in VS Code, Cursor, and JetBrains draw on that same allowance. The help center says there is no fixed message count. How much you can do depends on conversation length, the model, and the features you turn on.

The second part starts at the limit. Pro and Max both have a rolling five-hour window, and paid plans add a weekly limit. Max also has a separate weekly limit for Fable. When the window is gone you can wait, move up a plan, or turn on usage credits and continue at standard API rates. Decline that option and Claude Code stays inside the plan. Auto-reload in Console adds money only after you have already chosen API credits.

Published prices on 19 October

Claude plans

The table is the individual plans on the pricing page that day. Tax is extra. Free does not include Claude Code. The Enterprise FAQ lists $20 per seat per month plus usage at API rates, billed annually. The Team seat price was not on the comparison cards that day, so this article does not guess it.

Read the model rows on the same page. Opus, Sonnet, and Haiku are included on Pro and Max until you hit a limit. Fable on Pro is listed as usage credits, not as included allowance. On Max 5x and Max 20x, Fable is 50% of the weekly limit. Leave an agent on Fable under Pro and the first turn is already $10 per million input tokens and $50 per million output tokens.

Global list price, USD per million tokens, five-minute cache writes. Cache hits are from the 7 October rate card. US-only inference is 1.1x these input and output rates.

Opus 5.5 input and output are both 2x Sonnet 5.5. Fable 5.1 is 5x Sonnet. Haiku 5.5, while the prompt stays at or under 100K tokens, is one twentieth of Sonnet on both input and output. Cross 100K and the whole Haiku prompt is billed at $0.50 / $2.50, five times the short-prompt rate. The tail is not the only part that jumps.

Codex, Copilot, and Cursor

A coding agent is not only a Claude invoice. The seat prices below are the ones that still matched an official page on 19 October. Overage unit prices belong to each vendor’s checkout.

The three Pro prices are from About ChatGPT Pro tiers, updated that day: $100, $200, and $500, with Ultrafast only on $500. Plus at $20, including Codex, is from openai.com/codex. Free and Go can use Codex; Codex Cloud is not on those two plans. See Using Codex with your ChatGPT plan. Which GPT-6 tier Plus and Pro actually get is in the 16 October article. developers.openai.com/codex/pricing still shows a single Pro card at $200 and names GPT-5.4 and GPT-5.3-Codex. For the monthly fee, use the help center and checkout.

Copilot seats and credits are from the GitHub Copilot plans page. That page does not publish one overage rate for extra credits, so this article does not invent one. Cursor is from cursor.com/pricing that day. The page names X Subscription Corp as the seller. Plus is Grok only. From Premium up, third-party models are API list price on top of the monthly fee. The allowance refreshes weekly.

Why it costs more the more you use it

The seat price does not change. Three things stack: a longer context burns the included allowance sooner, every turn past the limit is list price, and list price jumps with the model.

One shared allowance

A long spec read in the afternoon and a repository edit in Claude Code that evening draw on the same five-hour window and the same weekly limit. The window rolls every five hours. The weekly limit resets at a fixed time on the account. The Max article says Pro is at least 5x Free, and Max 5x and 20x are 5x and 20x Pro, per five-hour window. That is not a token count you can multiply by the monthly fee. Anthropic does not publish how many tokens $20 buys. Hitting the wall one afternoon does not mean the month needs Max.

A longer context means fewer turns inside the same window. The pricing page says context goes up to 1M tokens and varies by model. Extra context adds no dollars while you are still inside the allowance. It pushes you to usage credits sooner. After that, each turn adds list price. That is the shape of “it costs more the more you use it”: marginal cost is $0, then marginal cost is the rate card.

Models a factor apart

One turn with 80,000 input tokens, 2,000 output tokens, and no cache hit, at global list price:

• Haiku 5.5, prompt still ≤100K. 80,000 / 1,000,000 × $0.10 + 2,000 / 1,000,000 × $0.50 = $0.009.
• Sonnet 5.5. 80,000 / 1,000,000 × $2 + 2,000 / 1,000,000 × $10 = $0.18. Exactly 20x the short Haiku prompt.
• Opus 5.5. 80,000 / 1,000,000 × $4 + 2,000 / 1,000,000 × $20 = $0.36. Twice Sonnet.
• Fable 5.1. 80,000 / 1,000,000 × $10 + 2,000 / 1,000,000 × $50 = $0.90. Five times Sonnet, and about 100 times short-prompt Haiku.

Those four lines show up on an invoice only on overage, or when Pro uses Fable. How to choose the model is in the September piece; use this table for the rates: Opus 5.5 and Claude Code.

Cache and long prompts

Writing an 80,000-token prefix into Sonnet’s five-minute cache costs 80,000 / 1,000,000 × $2.50 = $0.20. Sending it once as input costs $0.16. A one-off prompt is more expensive to cache than to send. On the second turn, if all 80,000 tokens hit: 80,000 / 1,000,000 × $0.10 = $0.008, plus $0.02 for 2,000 output tokens, $0.028 in total. Cache saves a repeated prefix. It does not save a one-time exploration.

Haiku’s jump is the prompt length, not a surcharge on the tail past 100K. 120,000 input plus 2,000 output: 120,000 / 1,000,000 × $0.50 + 2,000 / 1,000,000 × $2.50 = $0.065. The same token count at the ≤100K rates would be $0.013. The whole prompt becomes 5x. Getting the repository brief back under 100K matters more than trimming the question by a sentence.

An environment variable skips the plan

The help center says that if ANTHROPIC_API_KEY is set, Claude Code uses that key instead of the included usage on Pro, Max, Team, or Enterprise, and bills the API from the first turn. Run /logout, then /login with only the subscription account, and do not add Console credentials. After the limit, the API-credit prompt stays away. /status, or Settings → Usage on the web, shows what is left.

The other skip is inside Cursor. The Claude Code extension signed in with a Claude subscription draws on the Claude allowance. Picking a Claude model in Cursor’s own agent, on Premium and above, is API list price and is not included in the $30, $100, or $200. Both at once are two invoices.

The part of the month you can actually compute

The 800 turns below are an assumption: 40 turns on each of 20 workdays, 80,000 input tokens and 2,000 output tokens each, and all of them past the allowance. The warm-cache case assumes a 70% hit, so 56,000 tokens at the hit rate and 24,000 at the base input rate. This is not an Anthropic usage report, and $144 is not the average Pro bill. If those 800 turns stay inside the plan, overage is $0.

A warm turn is hit tokens × hit price + miss tokens × input price + output. Sonnet is 56,000 / 1,000,000 × $0.10 + 24,000 / 1,000,000 × $2 + $0.02 = $0.0736. Times 800 is $58.88. Warm Haiku is 56,000 / 1,000,000 × $0.01 + 24,000 / 1,000,000 × $0.10 + $0.001 = $0.00396. Times 800 is $3.168, written as $3.17 in the table.

Add the seat, still assuming all 800 turns are overage:

• Stay inside the allowance. Pro is $20, or $17 a month on the annual bill. Max 5x is $100. Max 20x is $200. Decline usage credits and there is no third line.
• Pro plus cold Sonnet overage. 20 + 144 = $164. Warm cache is 20 + 58.88 = $78.88.
• Pro plus cold Opus. 20 + 288 = $308.
• Pro plus cold Fable. 20 + 720 = $740. On Pro, Fable is outside the included allowance, so this line is easier to land on.
• Pro plus cold short-prompt Haiku. 20 + 7.20 = $27.20. Do not keep a refactor on Haiku just to save that $7.20.

Copilot Pro’s $15 of credits does not cover the $144 of cold Sonnet overage above. Cursor Premium’s $30 does not include Claude models. If those third-party tokens are the same Sonnet list price, the $30 seat still leaves $144 or $58.88 beside it. Grok usage stays on Cursor’s weekly pool and is a different line.

How to bring the development bill down

• Read /status before you upgrade. The five-hour window rolls. An afternoon that empties it may be back in the evening. The weekly limit is the signal that the month is short. Settings → Usage shows the next reset.
• For a hard ceiling, decline usage credits. Unset ANTHROPIC_API_KEY, sign in with only the subscription, and turn auto-reload off in Console. At the limit, Claude Code stops instead of continuing at list price.
• Default to Sonnet 5.5. Use it for ordinary edits. Hand renames and formatting to Haiku 5.5 with prompts at or under 100K tokens. Keep Opus 5.5 for the piece Sonnet keeps missing. Use Fable only for a long task that actually needs it, and remember that on Pro it is list price and on Max it is half the weekly limit.
• Cache a prefix that repeats. A one-time exploration costs more as a five-minute write than as a plain send. A stable repository brief and a fixed system prompt are what the hit rate is for.
• Keep Haiku prompts at or under 100K tokens. Past that, the whole prompt is 5x.
• Compact the context. Do not treat 1M as the default. A longer context spends the flat fee’s turns first, then walks you onto the rate card.
• Do not pay for Claude twice on the same repository. The Claude Code extension uses the subscription allowance. A Claude model selected inside Cursor is API list price. If $15 of credits covers completions and a light agent, Copilot Pro at $10 is cheaper than another Max seat. When it does not, add credits or one larger allowance. Do not turn both dials to the top.
• Keep the July price sheet and the June bill note in their months. The 9 July pricing note describes the plans as they were then. The 23 June drop from $800 to $150 is about the default model, context, and which billing mode you are in. The moves still apply. The rates are the 7 October card.

Which line is the Mac

None of the numbers above include a machine. When the command has to be xcodebuild, the simulator, or signing, an agent on Linux cannot do that step, and that is when a Mac joins the bill. On the plan page that day, M4 · 16GB / 256GB starts at $101.7 a month and 24GB / 512GB starts at $200.3 a month. Checkout is the plan page. Max 5x at $100 plus 16GB at $101.7 is $201.7 across two invoices. It is not the price of Claude.

API tokens, tool calls, and the cloud server for an agent are another line. See what an agent costs per month. Do not record the same machine once under the coding subscription and again under the server bill.

FAQ

How many tokens does Pro’s $20 include?

Anthropic does not publish that number. The help center says usage depends on conversation length, model, and features, and there is no fixed message count. Pro is at least 5x Free per five-hour window, plus a weekly limit, shared with chat on the web and with Claude Code. The $20 is the seat. Token rates appear after you turn on usage credits or switch to an API key.

Can Max’s $100 or $200 of API credits pay Claude Code overage?

No. The help center says Max’s monthly API credits cannot be used in Claude, Claude Code, or for usage credits. They are for the Claude Platform. Claude Code overage is a separate usage-credit switch, billed at standard API rates, or a wait until the five-hour window and the weekly limit reset.

What happens if ANTHROPIC_API_KEY is set?

Claude Code uses that key and bills API list price from the first turn. The subscription allowance is not used. To return to the plan, remove the variable, run claude logout, and sign in with the Pro or Max account only. Do not sign in to Console at the same time. If you want a stop at the limit, decline API credits and turn off Console auto-reload.

Is a month $20, $100, or several hundred dollars?

Inside the allowance it is the seat: Pro $20 (or $17 on the annual bill), Max $100 or $200. Several hundred dollars is overage. In this article’s example, 800 cold turns are $144 on Sonnet 5.5, $288 on Opus 5.5, and $720 on Fable 5.1. If those 800 turns stay inside the plan, overage is $0. Anthropic does not publish an average user’s token count. Do not file the example as your invoice.

Is Cursor or Copilot cheaper?

The seat can be. Copilot Pro is $10 with $15 of credits. Cursor Plus is $8 and includes only Grok. Cursor Premium is $30, and Claude and other third-party models are API list price on top. If you still run the same overage turns on a Claude model, the rate card does not fall because the shell changed. The savings come from staying inside an allowance, using a cheaper model, or dropping a second seat that does the same work.

Split the bill before you upgrade

Inside the allowance, Pro is $20 and Max is $100 or $200. Outside it, Sonnet 5.5, Opus 5.5, and Fable 5.1 are $2/$10, $4/$20, and $10/$50 per million tokens. Haiku 5.5 is $0.10/$0.50 at or under 100K tokens, and the whole prompt becomes $0.50/$2.50 past that. Max API credits do not pay for Claude Code. An API key in the environment makes the subscription allowance do nothing. Add a Mac only for the step that cannot leave macOS.

Further reading

• What an AI agent costs per month: API tokens, tool calls, and the cloud server
• Is GPT-6 worth the upgrade? Free, Plus, Pro, and Intelligent UI
• Opus 5.5 and Claude Code: price, tokens, and the coding workflow
• Field note: a Claude Code bill from $800 to $150