Guides / Pricing

Prompt caching in practice: what it saves on Claude, GPT-6, Gemini and DeepSeek

Updated November 5, 2026. Researched and drafted with AI help; every figure is checked against the linked source.

Most AI API bills are dominated by input tokens that the model has already seen: the same system prompt, the same codebase, the same contract, sent again on every request. Prompt caching lets the provider bill those repeated tokens at a much lower “cache read” price. This guide puts numbers on it, using the official list prices we track on the prices page.

What caching changes on the bill

A cached request is billed in three parts: fresh input tokens at the normal input price, cached tokens at the cache read price, and output at the normal output price. Caching never makes output cheaper. On some providers the first request that stores the cache also costs a bit more (a “cache write”); more on that below.

Here is the cache read price next to the normal input price for the models we track that publish one. Prices are per million tokens.

* Promotional price through December 31, 2026. From January 1, 2027 Google lists $1.50 input and $0.15 cache reads, still a 90% discount.

† DeepSeek prices shown are the peak-hour prices; off-peak input and output are half price. See DeepSeek V4 Pro for the hours.

‡ GPT-5.6 Sol is shown at the $5 / $30 that OpenAI’s main pricing page listed until it stopped showing per-token prices. OpenAI’s developer pricing page shows a promotional $4 / $20 with $0.40 cache reads, available until November 30, 2026.

§ Prompts up to 100,000 tokens. Above that, Claude Haiku 5.5 charges $0.50 input and $0.05 cache reads. Every example below stays under 100,000 tokens. Haiku 5.5, released October 7, 2026, succeeds Claude Haiku 4.5 ($1 input, $0.10 cache reads).

¶ Prompts under 200,000 tokens. From 200,000 tokens, xAI lists $4 input and $1 cached input for Grok 4.7.

Anthropic cut Claude Sonnet 5.5’s cache read price from $0.20 to $0.10 per million tokens on October 7, 2026, when it launched Haiku 5.5. The figures below use the new price.

The table covers the main models we compare. Mistral Large 3 is not in the table; Mistral’s API pricing docs list its cached input at $0.05 per million against $0.50 uncached, a 90% discount. Grok 4.7’s cache price isn’t on xAI’s models page, but xAI’s separate pricing page lists it at $0.50 per million.

The discount column hides something important: the standard is 90%, but a few models go further, and Grok 4.7 stays well short of it at 75%. Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol and both DeepSeek models discount cached tokens by 95% or more. Those are the models where caching changes the ranking most.

Example 1: a coding agent step

This is the same workload as our cost per 1,000 requests guide: each step sends 20,000 input tokens, of which 18,000 are the same context as the previous step, and gets 1,000 output tokens back. Here is what 1,000 steps cost without caching (all 20,000 input tokens at full price) and with caching (18,000 at the cache read price).

Caching cuts these agent bills by roughly two thirds. It can’t cut them further because the 1,000 output tokens per step are still billed in full, and on most of these models output is now the biggest part of the cached bill.

Notice what happens at the top. Claude Fable 5.1 and GPT-6 Astra have the same $10 / $50 headline price, so for the same token counts they cost the same without caching. With caching, Fable 5.1 is $13.50 cheaper per 1,000 steps because its cache read is $0.25 instead of $1. One caveat applies to every Claude comparison here: Anthropic says Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text than earlier Claude models, so the same prompt may not be the same number of tokens on Claude and GPT. Measure your own prompts before deciding. You can see the pair side by side on the Fable 5.1 vs GPT-6 Astra comparison.

Example 2: questions about one long document

Caching matters even more when the repeated part is big and the answers are short. Say you load a 50,000-token document (a contract, a manual, a codebase summary) and ask 1,000 questions about it, each question 200 tokens with a 400-token answer.

Here the saving is close to the cache discount itself, because almost all of each request is the cached document. With caching, Claude Opus 5.5 ($18.80) costs less than Claude Sonnet 5.5 without it ($104.40), and the two $2 / $10 models, GPT-6.1 Sol and Claude Sonnet 5.5, cost the same $9.40 now that both charge $0.10 for cache reads.

These figures leave out cache writes. Real usage also needs at least one full-price (or write-priced) request each time the cache expires, so treat them as the steady-state cost.

Cache writes: when caching costs more at first

Providers handle the first, cache-filling request differently:

• Anthropic charges a cache write of 1.25x the input price for a cache that lasts 5 minutes, or 2x for a 1-hour cache. Reads after that are billed at the cache read price.
• OpenAI lists a cache write price of $2.50 per million tokens for GPT-6.1 Sol (1.25x its $2 input) on its model page. OpenAI’s caching applies to prompts of 1,024 tokens or more.
• Google turns on implicit caching by default for Gemini. If you create an explicit cache yourself, Google also charges a storage fee per hour for keeping it.

The 1.25x write pays for itself on the first reuse. Take the 50,000-token document on Claude Sonnet 5.5: sending it twice without caching costs 2 × $0.10 = $0.20. Writing it to a 5-minute cache costs $0.125, and the second request reads it for $0.005, so $0.13 in total.

The 2x 1-hour write needs one more reuse. On Claude Sonnet 5.5, whose cache reads cost 0.05x the input price, two requests cost 2x + 0.05x = 2.05x the plain input price, slightly more than the 2x you’d pay without caching. From the third request on, the 1-hour cache is cheaper (2.1x vs 3x). Use it when requests come in more than 5 minutes apart but you expect at least three of them within the hour.

How to get the cache read price in practice

Caches match the beginning of your prompt. A request only gets the cheap rate for the part that is identical to what was cached, starting from the first token. In practice:

1. Put stable content first. System prompt, tool definitions, the document or codebase, then the changing part (the user’s question, the latest step) at the end.
2. Don’t put timestamps or request IDs at the top. One changed token early in the prompt means everything after it is billed as fresh input.
3. Keep requests close together. A 5-minute Anthropic cache that expires between steps has to be written again.
4. Check the usage fields. The major providers report how many tokens were read from cache in the API response. If that number is zero, your prompt prefix is changing.

Which model wins once you cache

Caching doesn’t just lower everyone’s bill by the same amount. It rewards the models with the deepest cache discounts:

• Among the $10 / $50 flagships, Claude Fable 5.1 is cheaper per token than GPT-6 Astra on cache-heavy workloads.
• Among the $2-input mid-tier models, GPT-6.1 Sol and Claude Sonnet 5.5 (both $2 / $10) now share a $0.10 cache read, half of Gemini 3.1 Pro’s $0.20 ($2 / $12). Our Sonnet 5.5 vs GPT-6.1 Sol guide covers the other differences.
• Grok 4.7 has the cheapest headline price of the $2-input models ($2 / $6) and is the cheapest of them without caching, but its 75% cache discount means it costs more than Claude Sonnet 5.5 or GPT-6.1 Sol once most input is cached ($19.00 vs $15.80 per 1,000 agent steps).
• DeepSeek has the steepest discounts of all (96.7% and 98%), and off-peak hours halve input and output prices on top.
• Gemini 3.8 Flash keeps its 90% discount after the promotion ends, but every price doubles on January 1, 2027.

Price is only half the decision. Before you move a workload to a cheaper model, check how people rate it today on its model page, and test it with your own prompts.

To run your own numbers, the cost calculator has a cache share setting that applies each model’s cache read price, and the compare pages show two models’ prices side by side. For the basics of tokens, batch and peak pricing, see AI API pricing explained.

Sources

1. Anthropic: Claude API pricing
2. Anthropic: Claude Haiku 5.5 (pricing and Sonnet 5.5 cache-read cut)
3. Anthropic: Prompt caching
4. OpenAI API pricing
5. OpenAI developer docs: pricing (GPT-6 models)
6. OpenAI developer docs: GPT-6.1 Sol model page
7. OpenAI developer docs: Prompt caching
8. Google: Gemini API pricing
9. Google: Gemini context caching
10. xAI docs: Pricing
11. DeepSeek API pricing
12. Mistral docs: API pricing