Appearance

Language

Mastering AI agents

AI Subscription or API: Which Costs Less in 2026? A Cost Calculator and Three Scenarios

ChatGPT, Claude, Copilot, Cursor, Devin and Google AI subscriptions side by side in US dollars, API bills worked out from token usage, and three scenarios (light chat, daily coding with an agent, a small app) run through the calculator. Prices from each vendor's official pricing page, rechecked 2026-11-18.

Kelvin Wang · · 9 min read

Key points

  1. A subscription is a fixed monthly price with each vendor's own usage limits; the API bills what you use. In each of the three scenarios, the most expensive model in the calculator bills exactly 100× the cheapest.
  2. Each vendor counts subscription limits in its own units (5-hour windows, weekly caps, credits) and none states them in tokens, so you can compare prices but not usage per dollar.
  3. The calculator's break-even: pick one plan and one model, and it gives the monthly tokens at which the API bill reaches the plan's price, to hold against your own usage.

It depends on how many tokens you use in a month, and on which model. The AI subscription vs API cost calculator lists every subscription’s monthly price and works out each model’s API bill from your token usage, and each scenario links to it with its numbers filled in.

All prices are in US dollars, as listed on each vendor’s official pricing page on November 16, 2026, and rechecked unchanged on November 18. Vendors can change prices at any time. The calculator prints the date of its own last price check, November 16, 2026, not the November 18 recheck. The calculator itself shows NT dollars, because I built it in Taiwan, but every plan and every API bill in it also shows the US$ amount.

What does each AI subscription cost a month?

Subscription prices US$ a month, billed monthly, as listed; checked 2026-11-18
  • ChatGPT Go 8
  • ChatGPT Plus 20
  • ChatGPT Pro $100 100
  • ChatGPT Pro $200 200
  • ChatGPT Pro $500 500
  • Claude Pro 20
  • Claude Max 100
  • Claude Ultra 200
  • Cursor Plus 8
  • Cursor Premium 30
  • Cursor Super 100
  • Cursor Ultra 200
  • Devin Pro 20
  • Devin Max 200
  • GitHub Copilot Pro 10
  • GitHub Copilot Pro+ 39
  • GitHub Copilot Max 100
  • Google AI Plus 4.99
  • Google AI Pro 19.99
  • Google AI Ultra 99.99
  • Google AI Ultra 20x 199.99

Source: chatgpt.com, claude.com, cursor.com, devin.ai, github.com, gemini.google

The prices come from each vendor’s own pricing page: ChatGPT 1, Claude 2, Cursor 3, Devin 4, GitHub Copilot 5 and Google AI 6. Easy things to miss:

  • Google sets its prices country by country. Its US page lists US$4.99, US$19.99, US$99.99 and US$199.99 6. Elsewhere the local price is not a straight conversion: Google AI Pro is £18.99 a month in the UK 7, €21.99 in Spain 8, ¥2,900 in Japan 9 and ₹1,950 in India 10. The other vendors’ pricing pages show US dollar prices.
  • Only Claude Pro shows an annual price. Claude Pro billed annually is US$200 up front, US$17 a month 2; none of the other plans listed shows an annual price on its page.
  • Claude Max and Ultra include API credits. Max comes with US$100 a month and Ultra with US$200 a month in API credits to build on the Claude Platform 2.
  • GitHub Copilot plans include GitHub AI Credits. Every plan comes with a monthly allowance of credits for Copilot features, at 1 credit = US$0.01 5; these are for use inside Copilot, not a general API.
  • Cursor’s plan allowances cover Grok models only. Plus offers Grok models only; Premium and above include generous Grok usage but no allowance for third-party models, which Cursor bills separately at API list price 3.
  • Tax may come on top, depending on your billing address. Claude and Cursor say on their pricing pages that their prices exclude tax 2 3, and Anthropic calculates tax from your billing address 11; the other pages, Google’s local-price pages included, don’t say whether tax is included. Some subscriptions can be billed in a currency other than US dollars: Anthropic’s invoice help says a new Pro, Max or Ultra subscription uses the currency, or currency options, offered to new subscribers in your region 12. If your card is charged in a foreign currency or by a merchant abroad, your bank may add a foreign-transaction fee.
  • Not listed: xAI SuperGrok and Perplexity Pro. Their official pages showed no readable price when the prices were checked.

How is an API bill worked out?

APIs bill by the token, priced per million tokens, with input (your prompt, documents, code) and output (the model’s answer) priced separately; output usually costs several times more. Many models charge less for input that hits the cache: send the same long prompt again and the repeat is billed at the cached price.

API prices are listed in US dollars, and Anthropic, Google and xAI say so outright 13 14 15. Google bills the Gemini API through Cloud Billing 16, which in many countries charges in local currency, such as yen in Japan, pounds in the UK, euros in Germany and rupees in India, converted from US dollars at rates set at the beginning of every month 17. Its tax page lists, for Google Cloud usage, 20% VAT in the UK, 10% GST in Australia and 10% JCT in Japan 18.

The calculator’s formula:

  1. Monthly input tokens × (1 − cached share) × input price
  2. plus monthly input tokens × cached share × cached price
  3. plus monthly output tokens × output price
  4. all divided by a million, which gives US dollars; the calculator then converts that to NT$ at its exchange rate and adds its Extras %

If you only know the length of a text: OpenAI puts a token at about 4 characters or 0.75 words of English text 19, and Google at about 4 characters for Gemini models 20. Anthropic says token counts depend on the tokenizer, the language and the type of content, and that Claude 4.7 and later models produce about 30% more tokens for the same text 13. The calculator also has a helper for Chinese text, whose ratio I measured with OpenAI’s tokenizers on 53 of my Traditional Chinese posts (about 0.66–0.90 tokens per character with the newer tokenizer and 0.92–1.25 with the older one, 10th–90th percentile); for an exact count in any language, use the vendor’s token-counting tool where it has one.

Scenario 1: light chat, 30 messages a day — what does it cost?

Say about 30 messages a day, each sent along with the earlier conversation: 50K input and 15K output tokens a day, 30 days, so 1.5M input and 450K output tokens a month. Open the calculator with these numbers.

Scenario 1: API bill per model US$ a month; 1.5M input and 450K output tokens, no cache
  • Claude Fable 5.2 30
  • Claude Haiku 5.5 0.38
  • Claude Opus 5.5 15
  • Claude Sonnet 5.5 7.5
  • DeepSeek Flash 0.99
  • DeepSeek V4 Pro 3.76
  • Gemini 3.5 Flash-Lite 1.58
  • Gemini 3.8 Flash 2.81
  • Gemini 4 Carbon 7.5
  • Gemini 4 Hydrogen 5.63
  • GPT-6 Astra 37.5
  • GPT-6 Luna 0.38
  • GPT-6.1 Sol 7.5
  • GPT-6.1 Terra 3.75
  • Grok 4.3 3
  • Grok 4.7 5.7
  • Grok Build 0.1 2.4

Source: claude.com, deepseek.com, google.com, openai.com, x.ai

For the same usage, the API bill runs from about US$0.38 to US$37.50 (exactly 100 times apart before rounding); the cheapest paid plan from each vendor costs US$4.99 to US$20 a month. At this volume the two ranges overlap, and which side costs less depends on the model.

Subscriptions also come with things the API doesn’t: web and mobile apps, file uploads, voice, search. With the API you bring your own interface or write code. The calculator compares money only; the features are yours to weigh.

Scenario 2: coding with an AI agent every day — what does the API cost?

A coding agent resends the project and the conversation at every step, so input is far larger than in chat, but most of it hits the cache. Say 20M input tokens per workday, 90% from cache, 200K output, 22 workdays a month. Open the calculator with these numbers.

The calculator uses each model’s price for shorter prompts, and a large project can push single requests past the size at which several vendors charge more: over 100,000 prompt tokens for Claude Haiku 5.5 13, over 272K input tokens for OpenAI’s flagship models 21, and 200K tokens or more for Grok 15. The calculator also leaves out cache-write charges: Anthropic and OpenAI charge 1.25 times the input price for tokens written to a short-lived cache 13 21, so heavily cached bills can come out higher. In this scenario, Claude Sonnet 5.5’s bill could rise by about US$22 if all of its uncached input is written to the 5-minute cache (more with longer-lived caches).

Scenario 2: API bill per model US$ a month; 440M input (90% cached) and 4.4M output tokens
  • Claude Fable 5.2 607
  • Claude Haiku 5.5 11
  • Claude Opus 5.5 343
  • Claude Sonnet 5.5 172
  • DeepSeek Flash 21
  • DeepSeek V4 Pro 93
  • Gemini 3.5 Flash-Lite 36
  • Gemini 3.8 Flash 79
  • Gemini 4 Carbon 172
  • Gemini 4 Hydrogen 129
  • GPT-6 Astra 1,056
  • GPT-6 Luna 11
  • GPT-6.1 Sol 172
  • GPT-6.1 Terra 86
  • Grok 4.3 145
  • Grok 4.7 312
  • Grok Build 0.1 132

Source: claude.com, deepseek.com, google.com, openai.com, x.ai

In this scenario the API bill runs from about US$11 to US$1,056. The assumptions are only an example; real usage varies a lot, and the same person can use several times more tokens on another project or when working differently. The most accurate input is the usage shown on your service’s dashboard.

Then use the calculator’s last part, “At what usage does the API cost as much as a plan?”. Pick a plan and a model, and it works out the monthly token count at which the API bill equals that plan’s price, using your input, output and cache mix. For the US$20 tier against every model:

What US$20 a month buys on the API Million tokens a month (input + output); scenario 2's mix: 99% input, 90% of it cached
  • Claude Fable 5.2 14.6
  • Claude Haiku 5.5 841.7
  • Claude Opus 5.5 25.9
  • Claude Sonnet 5.5 51.8
  • DeepSeek Flash 426.2
  • DeepSeek V4 Pro 95.6
  • Gemini 3.5 Flash-Lite 246.3
  • Gemini 3.8 Flash 112.2
  • Gemini 4 Carbon 51.8
  • Gemini 4 Hydrogen 69.1
  • GPT-6 Astra 8.4
  • GPT-6 Luna 841.7
  • GPT-6.1 Sol 51.8
  • GPT-6.1 Terra 103.6
  • Grok 4.3 61.2
  • Grok 4.7 28.5
  • Grok Build 0.1 67.3

Source: claude.com, deepseek.com, google.com, openai.com, x.ai

Scenario 2 uses about 444 million tokens a month; at its mix, one model passes US$20 within 10 million tokens, while two pass it only beyond 800 million.

Scenario 3: a small app calling an API — what does it cost a month?

Say a small tool gets 2,000 calls a day, each sending 1,200 tokens and getting 300 back, every day: 72M input and 18M output tokens a month. Open the calculator with these numbers.

Scenario 3: API bill per model US$ a month; 72M input and 18M output tokens, no cache
  • Claude Fable 5.2 1,296
  • Claude Haiku 5.5 16
  • Claude Opus 5.5 648
  • Claude Sonnet 5.5 324
  • DeepSeek Flash 43
  • DeepSeek V4 Pro 166
  • Gemini 3.5 Flash-Lite 67
  • Gemini 3.8 Flash 122
  • Gemini 4 Carbon 324
  • Gemini 4 Hydrogen 243
  • GPT-6 Astra 1,620
  • GPT-6 Luna 16
  • GPT-6.1 Sol 324
  • GPT-6.1 Terra 162
  • Grok 4.3 135
  • Grok 4.7 252
  • Grok Build 0.1 108

Source: claude.com, deepseek.com, google.com, openai.com, x.ai

An app calling a model from code is billed at API prices, so this scenario has no subscription side. The bill runs from about US$16 to US$1,620. If every call carries the same system prompt, enter a cached share and the input cost drops noticeably. DeepSeek charges half price off-peak, and the calculator uses its peak price; peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday except Chinese public holidays, and every other hour is off-peak 22.

How often are the calculator’s prices updated?

The prices live in one data file, each row with its source URL and the date it was checked, and the calculator prints the oldest date. Announced changes are noted next to their row: Gemini 4 Carbon’s price doubles from December 1, 2026, to US$4 input and US$20 output per million tokens, which takes its scenario 2 bill from US$171.60 to US$343.20 (the calculator shows the old price until its data file is updated), and Gemini 3.8 Flash’s price doubles from January 1, 2027 14. Some models on the vendors’ pages aren’t in the calculator yet: OpenAI’s GPT-6.2 Astra (US$8 input, US$40 output) and GPT-6.2 Sol (US$2 input, US$10 output) 21, and Google’s Gemini 4 Argon, priced like Carbon and doubling on the same date 14. If a date looks old or a price differs from the vendor’s page, the vendor’s page wins.

Frequently asked questions

Why doesn’t the calculator tell me which vendor to pick?

Price is one factor; models, interfaces, usage rules and data policies differ too, and only you know your usage. The calculator lists official prices side by side and works out the money for your usage, sorted by name by default; switch to “By price” to sort by cost.

Why does the calculator show NT dollars?

Because I built it in Taiwan. Each plan row lists its US$ price next to the NT$ one, and each API bill shows the US$ amount under the NT$ figure, so you can read US dollars directly. A few parts still assume Taiwan. The exchange-rate box (NT$ per US$1) works only from 1 to 100: a rate below 1 (euro, pound) or above 100 (yen, won) is quietly treated as 1 or 100, so the NT$ figures won’t match your currency. Google’s plans use their Taiwan NT$ prices (so their break-even figures differ slightly from the US$ list prices), and the Extras % box, where you can enter your tax rate, isn’t applied to them. The break-even card gives its token figure with an equivalent in Chinese characters, and shows its plan and API amounts in NT$ only.

Sources

  1. 1. Pricing (ChatGPT Learn) · OpenAI · checked 2026-11-18
  2. 2. Plans & pricing · Anthropic · checked 2026-11-18
  3. 3. Pricing · Cursor · checked 2026-11-18
  4. 4. Plans and pricing · Cognition · checked 2026-11-18
  5. 5. GitHub Copilot plans & pricing · GitHub · checked 2026-11-18
  6. 6. Google AI plans (United States) · Google · checked 2026-11-18
  7. 7. Google AI plans (United Kingdom) · Google · checked 2026-11-18
  8. 8. Google AI plans (Spain, in Spanish) · Google · checked 2026-11-18
  9. 9. Google AI plans (Japan, in Japanese) · Google · checked 2026-11-18
  10. 10. Google AI plans (India) · Google · checked 2026-11-18
  11. 11. Understanding your billing address and tax calculation · Anthropic · checked 2026-11-18
  12. 12. Understanding your Pro, Max, or Ultra plan invoices · Anthropic · checked 2026-11-18
  13. 13. Pricing (Claude Platform Docs) · Anthropic · checked 2026-11-18
  14. 14. Gemini Developer API pricing · Google · checked 2026-11-18
  15. 15. API Pricing (SpaceXSI Docs) · xAI · last updated 2026-10-10, checked 2026-11-18
  16. 16. Gemini API Additional Terms of Service · Google · checked 2026-11-18
  17. 17. Currencies available for Cloud Billing self-serve, auto-pay accounts · Google Cloud · last updated 2026-11-10
  18. 18. Taxes in your country (Cloud Billing) · Google Cloud · last updated 2026-11-10
  19. 19. Key concepts: tokens · OpenAI · checked 2026-11-18
  20. 20. Understand and count tokens · Google · checked 2026-11-18
  21. 21. API pricing · OpenAI · checked 2026-11-18
  22. 22. Models & pricing · DeepSeek · checked 2026-11-18

About the author

Kelvin Wang

Electrical engineer building tools with AI agents, sharing hands-on notes on engineering and investing.

More about me