Want to connect AI to your own code for summaries, translation, or customer support? Under this article’s assumptions, summaries cost about US$1.35 to US$27 a month, translation about US$0.72 to US$14.40, and customer support about US$4.65 to US$93. These are rough usage-fee estimates. For your own budget, recalculate with the token counts you actually measure.
How AI API Cost Is Calculated: Start With Tokens
An API is the channel that lets your code call an AI model. Its usage is billed separately from the usage limits of a chat subscription, and some subscriptions include API credits, explained below. API bills are charged by token, much like a mobile data plan is charged by data used. A token is the unit AI uses to measure the amount of text.
- The text you send in is called “input” and the text the model writes is called “output.” They are priced separately.
- Every provider lists prices as “US dollars per 1 million tokens.”
- For all four models in this article, the output price is 5 times the input price, so for the same number of tokens, output costs 5 times as much as input.
How Chinese Characters Convert to Tokens
When Google describes Gemini, it counts about 4 characters as 1 token, and says Chinese should be measured for real. Anthropic separately explains that models after Claude 4.7 tokenize text differently, producing about 30% more tokens for the same text, with the actual difference depending on the content.
This article roughly estimates “1 Chinese character is about 1 token,” which is easy to calculate in your head. It is not an official figure. For an exact count, measure the full request first (instructions, data, and question). Gemini count_tokens and Claude token counting both estimate input. For Claude, specify the target model, and measure again if you switch models. For actual output, run a small test and check the usage the API returns.
API Prices for Four Models
The prices below are the official standard-processing text-token prices for each API, without caching or batch discounts. For Gemini, we use Gemini 4 Hydrogen, the upgrade Google recommends for Gemini 3.8 Flash, which replaces the 3.8 Flash used in the older version of this article.
| Model | Input (per 1 million tokens) | Output (per 1 million tokens) |
|---|---|---|
| GPT-6 Luna (OpenAI) | US$0.10 | US$0.50 |
| Gemini 4 Hydrogen (Google) | US$1.50 | US$7.50 |
| Claude Haiku 5.5 (Anthropic) | US$0.10 | US$0.50 |
| Claude Sonnet 5.5 (Anthropic) | US$2 | US$10 |
Haiku 5.5’s price applies to a single prompt of up to 100,000 tokens. Above that, the whole request is charged US$0.50 for input and US$2.50 for output. The scenarios in this article are far below that.
AI API Cost Calculator: Three Scenarios for Summaries, Translation, and Support
Shared assumptions: a 30-day month, 1 Chinese character counted as 1 token, and no caching or batch discounts (both explained below).
- Summaries: Summarize 100 articles of 3,000 characters each per day, with a 300-character summary per article.
- Input: 100 articles × 3,000 characters × 30 days = 9 million tokens
- Output: 100 articles × 300 characters × 30 days = 0.9 million tokens
- Translation: Translate 20 articles of 2,000 characters each per day, with a translation the same length as the original. With the same arithmetic, that is 1.2 million input tokens and 1.2 million output tokens per month.
- Customer support: Answer 500 questions per day, with a 200-character reply each. This example assumes every question is sent together with 2,000 characters of support material, plus a 100-character question. With the same arithmetic, that is 31.5 million input tokens and 3 million output tokens per month.
Formula: monthly cost = (input tokens ÷ 1 million × input price) + (output tokens ÷ 1 million × output price). For example, summaries with Gemini 4 Hydrogen: 9 × 1.50 + 0.9 × 7.50 = US$20.25.
| Monthly cost (USD) | Summaries | Translation | Customer support |
|---|---|---|---|
| GPT-6 Luna | 1.35 | 0.72 | 4.65 |
| Claude Haiku 5.5 | 1.35 | 0.72 | 4.65 |
| Gemini 4 Hydrogen | 20.25 | 10.80 | 69.75 |
| Claude Sonnet 5.5 | 27.00 | 14.40 | 93.00 |
Within the same scenario, the most expensive and cheapest models differ by 20 times. To compare API prices, plug your own monthly input and output token counts into the prices in this table.
Price is not the only thing to look at when choosing a model. Anthropic’s official advice is to evaluate by task. You can start with Haiku 5.5 and upgrade if it is not capable enough. Compare reply quality on real data before you decide.
Plug In Your Own Scenario
Estimate how many requests you handle per day, and how much text each request sends in and asks the model to write. Take one real example and measure its input tokens with the target model, then run a small test of the output and check the usage the API returns. Multiply the tokens per request by the requests per day, then by 30 days, and apply the formula above.
Where Your Real Bill Will Differ From the Estimate
- Can be higher: Some models write out reasoning before they answer, called “thinking,” and thinking tokens are billed as output. Gemini 4 Hydrogen’s price includes thinking tokens, and its default thinking level is high, so recalculate with usage from a test run.
- Can be lower: Content you send repeatedly can use caching, where the platform remembers text you already sent and bills it at a lower price when you send it again. The read price is 10% of the standard input price for GPT-6 Luna and Haiku 5.5, and 5% for Sonnet 5.5 and Gemini 4 Hydrogen. The write price can be higher: GPT-6 Luna charges 1.25 times the standard input price. Claude charges 1.25 times for a 5-minute cache and 2 times for a 1-hour cache. With Claude, you must set cache_control in the request and resend the same opening content within the cache lifetime to get the cache price. If you only write to the cache and never reuse it, it may cost more. The minimum depends on token count, not character count: at least 4,096 for Gemini 4 Hydrogen, and at least 512 for Haiku 5.5 and Sonnet 5.5. Under this article’s assumption of 1 character per token, 2,000 characters of support material is enough for both Claude models but not for Gemini, so count tokens on real data first. Gemini 4 Hydrogen has only automatic caching, which is on by default. You get a discount only when repeated content hits the cache, and you cannot create a cache yourself.
- Can be lower: Non-urgent summaries and translation can use batch processing, which all three providers discount by 50%.
API or Monthly Subscription: Which to Choose
A chat subscription is paid monthly, while an API is priced by actual usage. If you want code to call a model, use the API. Claude Max, Ultra, and Team subscriptions include API credits: US$100 for Max and US$200 for Ultra, and Team credits are shared, with a cap of US$500. New subscribers can claim them after 7 days, and you need the plan and the necessary permissions in the Console (Anthropic’s developer dashboard). Credits can only offset usage on Claude’s official platform and do not roll over when they expire. Free, Pro, and Enterprise plans are not eligible. This article calculates usage fees before credits. For monthly subscription prices at each provider, see AI subscription price comparison.
Penchan’s Take
Monthly API cost is the input tokens and output tokens each multiplied by their price, then added together. First measure how many tokens your full request actually has, then apply the formula. If a support bot keeps sending the same material, consider caching, and count the write fee, cache lifetime, and hit rate together. For non-urgent work, consider batch processing.
Further Reading
- AI Subscription Pricing: Monthly Fees and Plans Compared
- Gemini 4 Hydrogen Explained
- Claude Haiku 5.5 Explained
- Claude Haiku vs GPT-6 Luna
- Google AI Studio Tutorial
- Claude Opus vs Sonnet
- The Sweet-Spot Trap of a Million Tokens
This article compares AI tool features and subscription plans for consumers. It is not securities or investment advice. Actual pricing is subject to the latest official announcements from each platform, and the information here may be out of date.