Claude Opus 4.7 API cost guide for long documents, long chats, and batch runs

Without prompt caching, the basic cost formula is:

Cost = input_tokens / 1,000,000 × 5
     + output_tokens / 1,000,000 × 25

With prompt caching, split reusable context from new context. The first write of reusable context is charged at the cache-write rate: $6.25/MTok for a 5-minute cache or $10/MTok for a 1-hour cache. Later cache hits or refreshes are charged at $0.50/MTok. New, uncached user input is still billed at the normal input rate, and model responses are still billed at the output rate.

One-off long-document analysis: input plus output

If you analyze a document once and do not ask follow-up questions, the budget is straightforward: the document, system prompt, and user question are input tokens; the model’s answer is output tokens.

Using the public Claude API rates:

For example, 300,000 input tokens plus 8,000 output tokens works out as:

300,000 / 1,000,000 × 5  = 1.50
8,000 / 1,000,000 × 25   = 0.20
Total                     = 1.70 USD

If you are migrating from an older model, do not blindly reuse old token estimates. Anthropic’s pricing documentation says Opus 4.7 uses a new tokenizer, and the token count for fixed text can increase by up to 35%.

So an earlier estimate of 300,000 input tokens may be safer to budget as 405,000 input tokens. With the same 8,000-token output:

405,000 / 1,000,000 × 5  = 2.025
8,000 / 1,000,000 × 25   = 0.20
Total                     ≈ 2.23 USD

Repeated Q&A over the same document: caching is the dividing line

The most common cost mistake in long-document products is billing the same large file as fresh input on every turn. If users will ask multiple questions about the same document, include prompt caching in the budget model from the start.

Assume:

• Document: 300,000 tokens
• New question per turn: 2,000 tokens
• Model answer per turn: 2,000 output tokens
• Cache type: 5-minute prompt cache

In this example, the first cached request is more expensive than sending the full document once. But by the second turn, caching is already cheaper overall:

No cache, two turns:       about 1.56 × 2 = 3.12 USD
5-minute cache, two turns: about 1.935 + 0.21 = 2.145 USD

The budget question is therefore not just document size. It is cache hit rate: will the same context really be reused, will the follow-up arrive before the cache expires, and how much new uncached material is added on each turn?

Long conversations: avoid rebilling the whole history every turn

Long chats follow the same cost logic as long documents. If your app sends a large conversation history back to the model on every turn, input costs accumulate quickly. Stable, reusable context should be evaluated for prompt caching.

Assume:

• Existing conversation history: 200,000 tokens
• New message per turn: 1,000 tokens
• Model output per turn: 2,000 tokens

Choosing between a 5-minute and 1-hour cache is a product-behavior decision, not just a price-table decision:

• If users usually ask follow-up questions within five minutes, start by modelling the 5-minute cache.
• If users often pause for more than five minutes but return within an hour, the 1-hour cache costs more upfront but may avoid repeated cache misses and rewrites.
• If return behavior is unpredictable, sample real traffic, measure cache hit rate, and then adjust the architecture.

Batch workloads: use public synchronous pricing as a conservative baseline

Batch workloads are common for offline analysis, data labelling, bulk summarisation, and large-scale classification. But unless you have confirmed the exact batch price for your account, contract, or platform endpoint, avoid putting an unverified discount into the official budget.

A conservative baseline is to estimate the job using the public synchronous Claude API price, then revise downward only after the applicable batch price is confirmed.

The same basic formula applies:

Total cost = total input tokens / 1,000,000 × 5
           + total output tokens / 1,000,000 × 25

Example: 10,000 tasks, each with 2,000 input tokens and 500 output tokens.

Total input  = 10,000 × 2,000 = 20,000,000 tokens
Total output = 10,000 × 500   = 5,000,000 tokens

Input cost   = 20 × 5  = 100 USD
Output cost  = 5 × 25  = 125 USD
Total        = 225 USD

That $225 estimate does not include any batch discount. If you later confirm a lower batch price, replace the unit prices in the formula with the actual rates.

Also check where you are buying access. If you are not calling Anthropic’s Claude API directly, the bill may differ. Third-party data from CloudPrice lists Opus 4.7 at $5 input and $25 output per MTok for Anthropic/global-style entries, while some AWS Bedrock regional codes are listed at $5.50 input and $27.50 output per MTok. Treat that as a prompt to verify your actual platform billing page, contract, and official documentation.

Add a buffer before you approve a budget

A token spreadsheet is usually too optimistic before launch. At minimum, build in room for three things:

1. Tokenizer migration risk. Opus 4.7’s new tokenizer can make fixed text count up to 35% higher in tokens.
2. Cache hit-rate uncertainty. Prompt caching lowers cost only when the same context is reused within the relevant cache window.
3. Real user behavior. Users may request longer outputs, retry jobs, upload larger files, or let conversation history grow beyond your original assumptions.

A practical, non-official budgeting buffer could look like this:

These multipliers are not Anthropic quotes. They are conservative planning ranges. Once the system is live, replace assumptions with real token logs, cache hit rates, and invoice data.

A quick monthly budget template

Without prompt caching:

Monthly cost ≈ daily requests × 30
             × (average input tokens / 1,000,000 × 5
                + average output tokens / 1,000,000 × 25)

With prompt caching, separate the cost buckets:

Monthly cost ≈ normal input cost
             + cache write cost
             + cache hit / refresh cost
             + output cost

Before implementation, define at least these variables:

Bottom line

For one-off long-document analysis, estimate with $5/MTok input plus $25/MTok output.

For repeated Q&A over the same large document, or long chats that carry a large history, model prompt caching before you finalise the architecture. In the 300,000-token document example above, a cached second turn costs about $0.21, compared with about $1.56 when the full document is resent each time.

For batch workloads, start with the synchronous public Claude API price unless your actual batch, cloud-platform, or contract pricing has been confirmed. And if you are moving older workloads to Opus 4.7, apply the possible 35% tokenizer increase first, then add an operating buffer. That will usually be closer to the invoice than a budget based on the headline token price alone.