How to save on prompt caching and batch requests: a calculator with live prices

Updated Nov 7, 2026

Two levers cut an AI API bill without changing the model. Prompt caching: when the start of your prompt repeats - a long system prompt, a document, tool definitions - many sellers charge the repeated part at a lower "cache read" price. Batch and flex: if the answer can come minutes or hours later, some sellers sell the same model cheaper as a batch or low-priority (flex) offer. The cache and batch calculator shows what both do for your volume, at every seller of the model.

How to use the calculator

1. Model id. Copy it from the model's page - it is under the name, for example anthropic/claude-sonnet-5-5.
2. Input and output tokens a month. Your monthly volume; the token counter on /tools helps if you only know requests and text length.
3. Share of input read from cache. The part of each prompt that repeats exactly from one request to the next. A chatbot with a 3,000-token system prompt and 1,000-token questions is about 75%; a RAG answer over fresh documents is near 0%.
4. The job can wait. Tick it when the results are not needed in real time - nightly reports, classification, evaluation runs. Then the seller's batch / flex offers are counted too.

The table lists every seller: the plain price, the price with the cache (where the seller publishes a cache-read price), and what you pay - the lower of the two.

Two examples from today's prices

• Claude Sonnet 5.5, 100M input and 10M output tokens a month, half of the input cached: $300 a month at Anthropic's plain price, $205 with the cache - about a third less. The cheapest seller's plain price is $150, and the cache takes it to $102.50.
• DeepSeek V4.1 Flash, the same volume, 80% cached, the job can wait: $10.77 a month at the cheapest plain real-time price, $5.87 with the cache and a batch offer - 45% less.

What the calculator does not count

• Writing to the cache. Some sellers charge extra once, when a prompt is first stored in the cache; with a high hit rate it is small, with a low one it can eat the saving.
• Cache lifetime. A cache entry expires after minutes of silence; rare requests may never hit it.
• Unpublished discounts. A batch discount the seller does not list as its own offer is not in the table.

Read next

• Prompt cache and batch calculator
• Cost calculator and other tools
• Model classes - to find a cheaper model of the same strength