How to Cut Claude API Costs by Up to 90% | Complete Guide to Batch Processing and Prompt Caching
Fable 5 is finally back—but the cost is a real concern.
“Claude API bills are higher than expected.” “I want the latest model, but I’m worried about cost.” Many teams embedding AI into business hit this wall.
In fact, Anthropic’s official API includes mechanisms from the start that can compress bills by half—or in some cases, close to one-tenth. They’re called batch processing and prompt caching.
This article organizes the latest model pricing as of July 2026, explains what these two mechanisms are, where they help, and how much you can actually save—with concrete calculation examples.
Claude API Pricing Refresher
Claude API pricing is pay-as-you-go, billed separately for input tokens and output tokens. A token is a unit of text processing; for Japanese, roughly 1 character ≈ 1–2 tokens is a useful rule of thumb.
Major model pricing as of July 2026 (per 1 million tokens):
| Model | Input | Output | Positioning |
|---|---|---|---|
| Claude Fable 5 | $10 | $50 | Top-tier Mythos class |
| Claude Opus 4.8 | $5 | $25 | Opus flagship |
| Claude Sonnet 5 | $2–3 | $10–15 | Standard model (intro pricing through end of August 2026) |
Worth noting: every model charges 5× for output tokens vs. input. So cost optimization comes down to two things: how to reduce wasted input and how to process the whole job—including output—at discounted rates.
Prompt caching helps with the former; batch processing helps with the latter.
What Is Batch Processing? How Bills Drop by Half
50% Off in Exchange for “Non-Urgent Processing”
Normal API calls are real-time: send a request, get a response in seconds to minutes. Batch processing (Message Batches API) bundles multiple requests and sends them to a dedicated endpoint for processing somewhere within 24 hours—asynchronously.
In exchange for “I don’t need an immediate reply,” both input and output tokens are half price. Fable 5 goes from $10/$50 to $5/$25; Opus 4.8 from $5/$25 to $2.50/$12.50.
Think of it like parcel pickup. One package or 100,000—contents can differ. The only promise is “everything arrives by the next day,” and shipping is half price.
Rules to Know
Batch processing has a few constraints.
The upper limit is up to 100,000 requests per batch, or 256MB of data—whichever comes first. There’s effectively no lower limit; even a single-request batch gets the discount.
Each request needs a unique ID. Results may not return in submission order, so you match inputs to outputs by ID. Each request in a batch is processed independently, so you can mix different models and prompts in one batch.
Most batches complete within an hour, but the guarantee is within 24 hours. Requests not processed after 24 hours aren’t billed. Results aren’t pushed automatically—you check status and download after completion, within 29 days of batch creation.
Good Use Cases for Batch Processing
Batch can’t support dialogue where you “ask the next question based on the answer,” so it fits one-way jobs that end with request → result. Typical cases:
Classification and tagging: auto-sorting tens of thousands of support emails, sentiment analysis on product reviews. Summarization and extraction: pulling contract terms and amounts from 1,000 contracts as structured data, nightly batch transcription summaries. Content generation: bulk product descriptions, SEO meta descriptions, bulk translation for multilingual sites. For developers, prompt evaluation and testing across hundreds of test cases is powerful.
The common thread: high volume, not urgent, fixed output format. Conversely, chatbots where users wait on screen for replies aren’t a fit.
What Is Prompt Caching? Cut Repeat Costs by ~90%
Stop Resending the Same Text Every Time
The API has no memory, so every request must resend “system prompt + reference materials + conversation history” in full. That repeated portion is billed at full input token rates each time—that’s normal.
Prompt caching stores the shared portion on the server once; later requests only reference the cache. Cache-hit input tokens cost roughly one-tenth of normal—about a 90% reduction.
Example: 100 Questions Against 100K Tokens of Documentation
Let’s put numbers on it. With Fable 5, load a 100,000-token manual and ask 100 questions.
Without cache: 100,000 tokens × 100 = 10 million input tokens—about $100 in input cost alone. With cache: the first call pays a slightly higher write-in rate, but the remaining 99 hits use cache read pricing, bringing the total down to about $11–12.
So the more you reuse long shared context, the bigger the win—powerful for long AI agent sessions, RAG over internal documents, and chat apps with extended conversations.
Caveats: TTL and What Gets Discounted
Cache has a TTL: 5 minutes by default, extendable to 1 hour. Unused cache expires and the next call starts from a write again. Also, discounts apply to input only—output tokens stay full price.
Combined Simulation: How Much Cheaper in Practice?
The two mechanisms can be used together—e.g., nightly batch document processing plus caching a shared system prompt.
Compare Opus 4.8 running 10,000 jobs: “50K tokens of shared instructions + 5K tokens of per-item data in, 3K tokens out.”
| Setup | Approximate cost |
|---|---|
| Normal API (no optimization) | ~$3,500 |
| Batch only | ~$1,750 |
| Batch + prompt caching | ~$700 |
Compared to doing nothing, roughly 80% savings is realistic. For high-volume teams, that gap can mean tens of thousands of yen per month.
Model Choice Is Part of Cost Strategy
One often-overlooked point: model selection itself.
Top-tier Fable 5 is exactly 2× Opus 4.8, but for complex long tasks, reports suggest fewer tokens and fewer retries to finish—so per-task cost can flip. For routine work, Sonnet 5 or Opus 4.8 usually delivers enough quality.
Practical recommendation: default to the standard model; switch to top tier only for hard tasks. Many teams compress bills significantly without dropping quality. Also set max_tokens (output cap) appropriately per call—a basic way to avoid wasting expensive output tokens.
Summary: Three Actions You Can Take Today
Claude API cost reduction doesn’t require special engineering. Three points:
- Move non-urgent bulk work to batch processing—input and output both become half price
- Put reused context in prompt caching—input cost drops by ~90%
- Match models to tasks—reserve top-tier models for when you truly need them
Pricing can change—always check Anthropic’s official pricing page before rollout. Try a small batch and cache setup first and see the difference on your bill.