LLM API Token Pricing Explained: What a Single Request Costs, Plus Caching and Long Context

How is a single request priced?

The LanoX pricing FAQ says: "Per million tokens is a pricing unit, so you pay for the tokens you use." Its own example uses made-up rates: "at hypothetical rates of $1 per million input tokens and $4 per million output tokens, a request using 2,000 input tokens and 500 output tokens would cost about $0.004. These rates are hypothetical examples, not current LanoX prices." (checked 2026-11-03)

The formula has 2 parts:

Output tokens often cost more than input tokens: 4 times more in this hypothetical example, and 5 times more for gpt-6.1-sol on both price lists below. That is why long answers cost more than long questions.

What does a real request cost?

Here is the same formula with real prices for gpt-6.1-sol, checked on 2026-11-03. The request uses 20,000 input tokens and 2,000 output tokens. This is our arithmetic from published base prices; long-context pricing is not included.

OpenAI also lists Batch prices for gpt-6.1-sol of $1.00 input and $5.00 output per 1M tokens (checked 2026-11-03). For every model's prices side by side, see the GPT API pricing page (coming soon) and the LLM API price gap tracker.

How does caching lower the cost?

When the same long prefix, such as a system prompt, repeats across requests, the repeated part can be billed at a cache-read price that is far below the normal input price. For gpt-6.1-sol on 2026-11-03:

Worked example (our arithmetic): if 15,000 of the 20,000 input tokens are read from cache, the LanoX cost of that later request is about 0.005 × $0.334 + 0.015 × $0.0167 + 0.002 × $1.67 ≈ $0.0053. Writing the cache in the first place is billed at the cache-write price, so caching pays off only when the same prefix is reused.

Anthropic's pricing page states the same idea for Claude: "A cache hit costs 10% of the standard input price" (checked 2026-11-03), with lower figures for some newer models listed on that page.

What about long context?

Long-context requests have tiered prices, so check the pricing page for the model you use. The examples in this article use base prices only.

Why can the same text cost different amounts on different models?

Because models split text into tokens differently, the same prompt can be a different number of tokens. Anthropic's pricing page says Claude 4.7 and later models "use a newer tokenizer" and "This tokenizer produces approximately 30% more tokens for the same text." (checked 2026-11-03)

So a lower price per token does not always mean a cheaper task. The LanoX pricing FAQ puts it plainly: "A lower input-token rate does not necessarily mean a lower cost for the finished task." Compare total cost per successful task, using the token counts each model reports.

Are images and video billed by token too?

Not always. The pricing FAQ says: "Image models may charge per image or by token usage... Video models may charge by generated duration, with different rates for resolution and generation mode." (checked 2026-11-03). On that date the catalog listed 12 image models and 12 video models. For one account that covers text, image, and video, see using 1 API account for text, images, and video (coming soon); for Grok's image and video prices, see the Grok API pricing page (coming soon).

Does a chat subscription include API credits?

Not necessarily. The pricing FAQ says: "A paid chat subscription does not necessarily include API credits, so check the account, billing, and access rules for each service." (checked 2026-11-03). API charges depend on usage; a subscription pays for features in 1 product. Check each service's own terms.

Who is per-token API pricing not a good fit for?

• People who need the same bill every month. API charges follow usage, so a busy month costs more.
• Anyone who will compare only the price per token. Anthropic says its newer tokenizer produces about 30% more tokens for the same text, and output can cost 5 times input.
• Workloads that are mostly images or video. Those can be billed per image or per second, so token math does not apply.
• Requests that regularly hit long-context tiers. Use the tiered prices on the pricing page, not the base-price examples here.

How do you check what you were actually charged?

Do not rely on estimates alone. The LanoX docs say to "Check the requested model, actual result, usage record, and final charge in the console" (checked 2026-11-03), which is 4 records per call. Send a small batch of real requests from your app, read the token counts and charges, and multiply up.