AI Token Pricing and API Costs AI-pedia

AI token pricing is the system that providers of large language models (LLMs) use to charge customers for access to their models through an application programming interface (API). An API lets one computer program send requests to another. In this case, a developer's software sends text, images or other data to a model hosted by the provider and receives the model's response. Instead of charging per request or per minute, almost all providers charge by the number of tokens processed. A token is a small chunk of text, typically part of a word.

Prices are usually quoted in US dollars per million tokens. Tokens sent to the model (the input or prompt) and tokens generated by the model (the output or completion) are counted and priced separately. Output tokens nearly always cost more. Prices vary widely by model. In the mid-2020s, the cheapest models cost a few cents per million input tokens. The most capable models cost tens of dollars per million output tokens, a difference of several hundred times or more.

Token prices have fallen sharply since commercial LLM APIs first appeared around 2020. Prices change often, and providers regularly release new models, retire old ones and adjust rates. Any specific figure in this article should be read as an example from a particular date, not as a current price. Current prices are published on each provider's pricing page.

What a token is

Tokenization

Language models do not read text as letters or whole words. Before text reaches the model, a program called a tokenizer splits it into tokens drawn from a fixed vocabulary. This vocabulary typically holds tens of thousands to a few hundred thousand entries. Common words are often a single token. Rarer words are split into several pieces, and punctuation marks and spaces are usually parts of tokens too.

Most modern tokenizers use a method called byte-pair encoding (BPE) or a close variant. BPE starts from individual characters or bytes and repeatedly merges the most frequent adjacent pairs found in a large body of training text. Common sequences therefore end up as single tokens. OpenAI publishes an open-source tokenizer library called tiktoken, and several providers offer token-counting tools or API endpoints so developers can measure text before sending it.

Rules of thumb

For ordinary English text, a widely used rule of thumb from OpenAI's documentation is that one token is about four characters, or about three-quarters of a word. By this rule, 100 tokens is roughly 75 words, and one million tokens is roughly 750,000 words. That is several long novels' worth of text. These are approximations only. The real count depends on the tokenizer and the content:

• Computer code, numbers, unusual spellings and text with many symbols tend to use more tokens per word.
• Different providers use different tokenizers, so the same text can produce different token counts on different models. This makes direct price comparisons less exact than they appear.
• A provider may change its tokenizer between model generations, which changes the count for identical text.

Language differences

Tokenizers are trained mostly on English and other widely represented languages, so text in many other languages is split into more tokens for the same meaning. Research published in 2023, including a study by Aleksandar Petrov and colleagues, found that the same content can need several times as many tokens in some languages as in English. Because billing is per token, users writing in those languages may pay more for equivalent work. Their prompts also use up more of the model's context window, which is the maximum amount of text the model can consider at once.

Non-text inputs

Models that accept images, audio or video also convert these inputs into tokens for billing. Images are usually charged by size and resolution, with larger or more detailed images counted as more tokens. Audio is often charged per second or per minute of sound, or converted to an equivalent token count. Some providers set separate, higher rates for audio tokens than for text tokens.

How pricing is structured

Input and output tokens

The basic price of a request is calculated from two numbers:

• Input tokens: everything sent to the model. This includes the user's message, any instructions (often called the system prompt), earlier conversation turns, attached documents and descriptions of available tools.
• Output tokens: everything the model generates in reply.

Output is priced higher because generating text costs the provider more computing work per token. A model can process the whole input in parallel in one pass. Output must be produced one token at a time, and each new token requires another pass through the model. Across major providers, output prices have commonly been about 3 to 8 times the input price for the same model. Some examples at list prices:

• Anthropic's Claude 3 family (March 2024) charged 5 times as much for output as input.
• OpenAI's GPT-4o (August 2024 version) charged 4 times as much.
• Google's Gemini 2.5 Pro charged 8 times as much for shorter prompts.

Model tiers

Most providers sell several models at different prices. They can be grouped loosely into tiers:

• Light or small models: the fastest and cheapest. They suit simple, high-volume tasks such as classification, short summaries or data extraction. Examples have included Claude Haiku, GPT-4o mini and the Gemini Flash-Lite models.
• Mid-range models: a balance of capability and cost, suitable for most general work.
• Flagship or frontier models: the most capable and most expensive models. They are used for difficult reasoning, complex coding or high-stakes writing.

The price gap between tiers is large. When Anthropic launched its Claude 3 family in March 2024, the list prices per million tokens were:

Model	Input ($ per million)	Output ($ per million)
Claude 3 Haiku	0.25	1.25
Claude 3 Sonnet	3.00	15.00
Claude 3 Opus	15.00	75.00

The top model in this family cost 60 times as much as the bottom one.

Context-length pricing

Some providers charge a higher rate when a single request is very long. For example, Google priced Gemini 2.5 Pro at $1.25 input and $10 output per million tokens for prompts up to 200,000 tokens. For longer prompts the rate rose to $2.50 input and $15 output. The higher rate reflects the extra computing and memory cost of processing very long inputs.

Reasoning tokens

Reasoning models generate an internal chain of intermediate steps before giving a final answer. OpenAI's o1, released in 2024, was an early example. These intermediate "thinking" or "reasoning" tokens are generally billed as output tokens, even when the provider does not show them in full to the user. As a result, a short visible answer from a reasoning model can still use thousands of billed output tokens. Some providers let developers set a maximum "thinking budget" to limit this.

Other charges

Besides basic token prices, providers may charge for:

• Tool use: built-in web search, code execution or file search is sometimes billed per call or per session, in addition to the tokens involved.
• Embeddings: separate, much cheaper models that turn text into lists of numbers used for search and retrieval. These are usually priced per million input tokens, with no output charge.
• Fine-tuning: further training of a model on a customer's own data. This is typically charged per training token, and fine-tuned models may then have higher usage prices than the base model.
• Image, audio and video generation: often priced per image, per second of audio or per second of video rather than per token.
• Storage: some caching and file features charge for keeping data on the provider's systems over time.

Discounts and ways to reduce cost

Batch processing

Major providers, including OpenAI, Anthropic and Google, offer batch interfaces. Customers submit many requests together and accept that results may take up to a set period, commonly 24 hours. In return, they get a substantial discount, often 50% off normal prices. Batch processing suits work that is not urgent, such as processing large document collections or running evaluations. It lets the provider schedule the work when its computers would otherwise be idle.

Prompt caching

Many applications send the same long beginning of a prompt many times, for example a fixed set of instructions or a large reference document. Prompt caching lets the provider store the processed form of that repeated portion and reuse it. Repeat reads are then billed at a reduced rate. The details differ by provider:

• Anthropic introduced caching in 2024. Writing content to the cache costs somewhat more than normal input, and reading from the cache costs a small fraction of the normal input price, around one-tenth. Cached content expires after a short period unless it is used again.
• OpenAI introduced automatic caching in 2024 for prompts above a minimum length. Cached input tokens are billed at a discount, with no action needed from the developer.
• Google offers context caching for Gemini models. Discounted reads are combined with a charge for storage time.

For applications that repeat large prompts often, caching can cut input costs by a large proportion.

Service tiers and contracts

Some providers offer a cheaper, slower "flex" tier with lower priority. Others offer a more expensive "priority" tier with guaranteed speed or capacity. Large customers can also negotiate committed-use contracts, reserved capacity or enterprise discounts. These rates are generally not published.

Free tiers and credits

Some providers give limited free use, either as starting credits or as a free tier with strict rate limits. Google's AI Studio, for example, has offered free access to some Gemini models with usage caps. Free tiers may come with different data-use terms from paid tiers. For instance, the provider may be allowed to use submitted data to improve its products.

Practical techniques

Developers commonly lower costs by:

• sending simple tasks to cheaper models and only harder tasks to expensive ones (sometimes called model routing or cascading);
• shortening system prompts and removing unneeded context;
• summarizing or trimming long conversation histories;
• setting a maximum output length;
• using retrieval to send only the relevant parts of large documents rather than the whole document.

Calculating costs

The basic formula

The cost of a request, or of a total workload, is:

cost = (input tokens ÷ 1,000,000 × input price)
     + (output tokens ÷ 1,000,000 × output price)

Cached tokens, reasoning tokens and extra charges are added where they apply. Because output usually costs more, the ratio of input to output strongly affects the total. A workload that reads long documents and writes short answers costs much less than one with the same total tokens that generates long text.

Worked example: 10 million tokens

The table below shows the cost of 10 million tokens at several sets of example prices. These prices are illustrative and based on past list prices. They are not current quotes. Three splits are shown: all input, a mix of 8 million input and 2 million output, and all output.

Example price (input / output per million)	Comparable to	All input	8M in + 2M out	All output
$0.10 / $0.40	Light tier (e.g. Gemini 2.5 Flash-Lite list price)	$1.00	$1.60	$4.00
$0.15 / $0.60	Light/mid tier (e.g. GPT-4o mini at launch)	$1.50	$2.40	$6.00
$1.25 / $10.00	Flagship (e.g. Gemini 2.5 Pro, shorter prompts)	$12.50	$30.00	$100.00
$3.00 / $15.00	Flagship/mid (e.g. Claude 3.5 Sonnet)	$30.00	$54.00	$150.00
$15.00 / $75.00	Premium (e.g. Claude 3 Opus)	$150.00	$270.00	$750.00

The example shows two points. First, the same 10 million tokens can cost anywhere from about a dollar to hundreds of dollars depending on the model. Second, within one model, the input/output split can change the total by a factor of several times. A batch discount of 50% would halve each figure.

Published estimates of costs for named models should be checked against the provider's own pricing page. Model names and prices change often, and secondary sources sometimes give outdated or incorrect figures. For example, some sources have quoted $0.075 input and $0.30 output per million tokens for Gemini 2.5 Flash-Lite. Those figures match earlier Google "Flash" and "Flash-Lite" models. Gemini 2.5 Flash-Lite's general-availability list price was $0.10 input and $0.40 output.

Hidden sources of token use

Token use is often higher than the visible conversation suggests:

• Conversation history: chat APIs are generally stateless, meaning the model does not remember earlier turns by itself. To continue a conversation, the application resends the earlier messages as input with each new request. A long conversation therefore costs more per turn as it grows, and the total cost of the whole conversation grows faster than its length.
• System prompts and tool definitions: these are included in every request and can be thousands of tokens long.
• Agents: software "agents" that let a model take many steps, such as searching, reading files and running code, can make dozens or hundreds of model calls for a single task.
• Reasoning tokens: as described above, these can greatly exceed the visible answer.
• Retries: requests that fail or produce unusable output and are run again are usually billed each time.

History of prices

Early commercial APIs

OpenAI opened a commercial API for its GPT-3 models in 2020. At first it priced usage per 1,000 tokens rather than per million. The largest GPT-3 model, "davinci", cost $0.06 per 1,000 tokens, equal to $60 per million, with the same rate for input and output. OpenAI cut this to $0.02 per 1,000 tokens in 2022.

The 2023–2025 period

From 2023 onward, competition grew and providers released many models. Several pricing changes became standard: separate input and output rates, the move to quoting prices per million tokens, and large price cuts with each model generation. Some notable list prices at launch:

Model	Provider	Launch	Input ($/million)	Output ($/million)
gpt-3.5-turbo	OpenAI	March 2023	2.00	2.00
GPT-4 (8K context)	OpenAI	March 2023	30.00	60.00
GPT-4 (32K context)	OpenAI	March 2023	60.00	120.00
GPT-4 Turbo	OpenAI	November 2023	10.00	30.00
Claude 3 Opus	Anthropic	March 2024	15.00	75.00
GPT-4o	OpenAI	May 2024	5.00	15.00
GPT-4o mini	OpenAI	July 2024	0.15	0.60
o1	OpenAI	2024	15.00	60.00
Gemini 2.5 Pro (≤200K prompt)	Google	2025	1.25	10.00

OpenAI also lowered GPT-4o's price to $2.50 input and $10 output in a later 2024 version. In early 2025, the Chinese company DeepSeek drew wide attention by offering its R1 reasoning model at prices far below those of comparable Western reasoning models at the time. This added to downward pressure on prices across the industry.

The overall trend

Between 2023 and 2025, the price of reaching a given level of performance fell by a large factor. GPT-4 in March 2023 charged $30 per million input tokens. Little over a year later, GPT-4o mini charged $0.15 per million input tokens, 200 times less, and its makers reported that it matched or beat the older GPT-4 on several common benchmarks. A widely cited 2024 analysis by the venture capital firm Andreessen Horowitz called this trend "LLMflation". It estimated that the cost of a fixed level of capability was falling by roughly ten times per year. Other researchers have reported broadly similar steep declines, though estimates vary with the benchmark and time period used.

The price of the most capable models has not fallen in the same way. New top-tier models are often launched at high prices, and premium tiers have remained expensive. In practice, the cheapest price for "good enough" performance has fallen dramatically, while the price of the leading edge has varied.

Why prices have fallen

Several factors have been cited for the decline in price per token:

• Better hardware: newer graphics processing units (GPUs) and specialized AI chips process more tokens per second for each dollar spent and each unit of electricity.
• Software efficiency: improvements in serving software, such as batching many users' requests together, better memory management and faster attention algorithms, increase how many tokens a given machine can handle.
• Smaller, better models: through distillation (training a small model to imitate a larger one) and improved training data, smaller models now match older large ones. Smaller models are cheaper to run.
• Quantization: storing model numbers at lower precision reduces memory and computing needs, with modest effects on quality.
• Mixture-of-experts architectures: these models activate only part of their parameters for each token, lowering the cost per token compared with a dense model of similar total size.
• Competition: many companies now offer capable models, and open-weight models (models whose trained parameters are published for anyone to run) set a price floor that hosted services must compete with.

Subscriptions compared with API pricing

Consumer chat products such as ChatGPT, Claude and Gemini are usually sold as monthly subscriptions, not per token. Their most common paid plans have been around $20 per month, with more expensive plans for heavy users. Subscriptions usually include usage limits that are described in messages or time windows rather than in exact token counts.

API pricing differs in several ways:

• it charges exactly for usage, with no fixed monthly fee in most cases;
• it is intended for building software, not for direct chatting;
• it usually requires a payment method and an account with usage tracking;
• it can be cheaper or more expensive than a subscription, depending on how much is used.

Light users of an API may pay only a few cents or dollars per month. Heavy users, especially those running agents or reasoning models, can spend far more than a subscription costs.

Third-party hosting and self-hosting

Cloud platforms and aggregators

Models are also sold through cloud platforms such as Amazon Bedrock, Microsoft Azure and Google Cloud Vertex AI. These usually charge similar per-token prices to the model makers, sometimes with regional differences or extra enterprise features. Aggregator services such as OpenRouter offer access to many models through a single interface and pass on per-token prices from various hosts. Open-weight models such as Meta's Llama, Mistral's models, Alibaba's Qwen and DeepSeek's models are offered by many competing hosts, often at low per-token prices.

Running models directly

An organization can also run open-weight models on its own hardware or on rented cloud GPUs. In that case it pays for hardware or GPU time (often billed per hour) rather than per token. Self-hosting can be cheaper at very high, steady volumes and gives more control over data. It also requires technical expertise, and the hardware still costs money when idle. Whether it saves money depends heavily on how fully the hardware is used.

Rate limits and usage tiers

Providers limit how much each customer can use in a given time to protect their systems and manage capacity. Limits are usually set as:

• requests per minute (RPM);
• tokens per minute (TPM), sometimes split into input and output;
• daily or monthly spending caps.

Many providers place accounts in usage tiers that rise as the customer spends more or keeps an account longer. Higher tiers allow higher limits. Developers can also set their own spending limits and alerts to avoid unexpected bills.

Debates and open questions

Are prices sustainable?

Commentators have debated whether current prices cover providers' real costs. Training frontier models costs very large sums, and serving them requires expensive data centres. Some argue that low prices are partly subsidized by investor funding to win market share. Others argue that efficiency gains make low prices profitable at scale. Providers generally do not publish detailed figures on the cost of serving each token, so this question remains open.

Cheaper tokens, more spending

Some observers have pointed to the Jevons paradox: the observation that when a resource becomes cheaper to use, total use can rise so much that total spending increases. Falling token prices have been accompanied by much heavier use, including long-context applications, agents and reasoning models that consume many more tokens per task. In those cases the cost per task may stay level or even rise while the price per token falls.

Cost per token compared with cost per task

Because models differ in how many tokens they use, the per-token price alone can mislead. A model with a lower per-token price may use more reasoning tokens or need more retries, making it more expensive for a given job. Some benchmark organizations now report the total cost to complete a test set, not just token prices.

Transparency

Billing for hidden reasoning tokens has raised questions about transparency. Users pay for tokens they cannot fully inspect. Differences between tokenizers also make it hard to compare prices across providers precisely, because "one million tokens" does not mean the same amount of text everywhere.

Fairness across languages

As described above, speakers of languages that tokenize less efficiently pay more for equivalent use. Researchers have suggested improving tokenizers for more languages or pricing by characters or meaning instead. Neither approach has become standard.

Environmental cost

Token prices do not directly reflect the energy and water used by data centres. There is ongoing discussion of the environmental footprint of AI use and how it relates to the volume of tokens processed. Published estimates of energy per query vary widely and are disputed.

See also

• Large language model
• Tokenization (natural language processing)
• Byte-pair encoding
• Application programming interface
• Context window
• Prompt engineering
• Retrieval-augmented generation
• Reasoning model
• Open-weight model
• Cloud computing pricing
• Graphics processing unit
• Jevons paradox