LLM Pricing Cheat Sheet: Compare Costs & Usage Tiers

Last update - 20 August 2026

Latest updates: Current flagship, value, embedding, and fine-tuning prices from official provider pages

How much does it cost for you to integrate your application with a large language model? Here's an article that will show the rough relative costs of using different models for chat / completion models, costs for embedding and for making use of a fine-tuned model. I will keep this page updated, so bookmark this and use this as a reference guide whenever you're working on your next cool idea. Happy building!

Additionally, if you have information on other models OR if you have additional pricing inputs to share, do write in the comments.

A quick intro before moving ahead with the pricing comparison

Interest and usage of large language models have exploded. Most teams have already tried general-purpose AI assistants, and the next value unlock is to apply a model to a specific use case and application. The 2 key questions to answer for you are:

1. Which model should I choose?

2. How much will it cost me / my organization?

While I cannot answer question 1 in this article (the short answer is: it depends), we can certainly look into question 2 to understand how much it costs, in order for us to make a better business case for the development of a custom app. If you're looking for some help with assessing which model to use, its solution design or development of your LLM based use case, fill up this form and let's talk!

Tokens vs Characters vs Everything else

The main challenge in comparing providers is that the headline token rate is no longer the full price. Providers can charge different rates for input, cached input, output, long context, tools, search, regions, data residency, and service tiers. The tables below normalize text prices to US dollars per million tokens where possible, and call out important extra charges.

Some definitions:

1. Session - I created the concept of a "session" to make it easier for us to visualize costs for an application we're building out. The definition of 1 session is: A session between a human and your bot consisting of 100 words x 5 characters each = 500 characters = 125 tokens (in english language). Assume 25% of the tokens are consumed by a human and 75% is consumed by your bot generating outputs.

2. Page - I like to think about embeddings in terms of pages rather than tokens. In this sheet, I've assumed 1 page = 500 words x 5 characters = 625 tokens per page

Pricing

So let's get into it!

A. Usage of General Prompt / Completion or Chat models

Let's first look at costs for all completion and chat models, the ones that we would use for most often: "ChatGPT for my App", chatbots, knowledge retrieval bots (+ add costs of embeddings to this)

Prices are in USD and were verified on 20 August 2026. The table now contains 88 priced model rows. Rates are standard direct-API or named global rates unless a note says otherwise. Batch, flex, priority, data-residency, regional, long-context, tool-call, search, and image/audio token charges can change the total. AWS Bedrock and Azure prices can also differ by region and endpoint, so check their calculators before you commit to a budget.

Perplexity currently publishes four first-party Sonar models. DeepSeek currently publishes two direct API models. Cohere also lists Command A+ and Command R7B, but it does not publish a comparable self-serve token rate for them, so they are not shown as priced rows.

Official sources: OpenAI pricing and model catalog, Anthropic pricing, Google Vertex AI pricing, xAI pricing, Mistral API pricing, Cohere pricing, Perplexity pricing, Groq models, DeepSeek pricing, and AWS Bedrock pricing.

1. For a mixed chat workload, calculate input and output separately. The earlier 25% input / 75% output session split remains a useful rough estimate.

2. A session of 125 tokens is only a planning unit. System prompts, retrieved context, reasoning tokens, cached prompts, and tool results can make real requests much larger.

B. Embedding Costs

What does it cost for you to embed your knowledge base using one of the embedding models available? Let's look at 3 scenarios - a knowledge base of 10,000 pages, 100,000 pages and 1 million pages.

1. This estimate uses 625 tokens per page, as defined above. It covers embedding API use only.

2. Vector database, storage, retrieval, reranking, network, and re-embedding costs are not included.

3. Google lists separate prices for image, video, and audio inputs to its multimodal embedding model.

C. Cost of Fine-tuning a model + usage:

Fine-tuning prices are now harder to compare directly. Some providers use training-token prices, some use training hours, and others provide custom enterprise quotes. OpenAI is also winding down access to its fine-tuning platform. Use the table below as a planning reference, not as a complete project quote.

Training cost is not the full cost of a tuned model. Add dataset preparation, evaluation, repeated experiments, model storage, and production inference. Check the provider page again before you start a training job.

Pricing scenario: a current monthly chat estimate

For a simple comparison, assume 100,000 sessions per month and 125 tokens per session. With the earlier 25% input / 75% output split, that is 3.125 million input tokens and 9.375 million output tokens per month.

These figures do not include system prompts, retrieved documents, reasoning tokens, tool calls, search fees, caching, or application infrastructure. In a retrieval-augmented generation system, embedding 100,000 pages once would add about $1.25 with OpenAI text-embedding-3-small or $9.38 with Google Gemini Embedding, before vector database costs.

Fine-tuning and retrieval are not interchangeable. Retrieval is usually the better fit when facts change often or citations matter. Fine-tuning is usually a better fit when you need repeatable behavior, style, or task-specific patterns. Many production systems use both.

Did you enjoy this article? Have I missed something in the pricing calculations? Drop a note below and get the discussion started!