Anthropic publishes two discount mechanisms and a batch endpoint that, used together, typically cut a Claude bill more than any prompt-level trimming can on its own.
Cache reads are 90% cheaper — after you pay a write premium first
Writing a new entry to Anthropic's prompt cache costs a premium over a standard input token — roughly 1.25x for the shorter, 5-minute default cache window, or 2x for the longer, 1-hour window — but every subsequent read of that same cached content costs only around a tenth of the standard input rate. That trade only pays off with genuine reuse: a system prompt, tool definitions, or reference document that gets hit repeatedly within the cache window earns the discount back many times over, while content that changes on every call just pays the write premium with no read savings to offset it. The minimum prompt length to qualify for caching varies by model, roughly in the 500–4,000 token range, so very short prompts won't be eligible regardless of TTL choice.
The Message Batches API is 50% off, and it stacks with caching
A flat 50% discount applies to both input and output tokens on batch requests, with most batches completing within about an hour and a guaranteed same-day turnaround. Anthropic's own documentation is explicit that prompt caching discounts apply on top of the batch discount, so cache-eligible, batch-eligible work — bulk classification, evals, offline document processing — gets both cuts together. The trade is the same as any batch endpoint: no streaming, no real-time response, and a cap of 100,000 requests or 256 MB per batch.
Extended thinking tokens are billed as output — set a budget
Like reasoning tokens on other providers, Claude's extended-thinking tokens count as output tokens — the more expensive half of a request — even though they aren't part of the visible answer. A prompt that triggers deeper thinking than the task actually needs inflates cost without changing what the response looks like. Setting a deliberate thinking budget rather than leaving it uncapped keeps this bounded instead of leaving it entirely to the model's judgment on a given prompt.
Match the model tier to the task, not the other way around
Anthropic's smaller, faster Claude tiers cost a fraction per token of its most capable model and handle summarization, classification, and other well-scoped tasks without needing frontier-level reasoning. Reserving the top-tier model for prompts that genuinely require it, and routing everything else to a cheaper tier, is usually a bigger cost lever than any single caching or batching change on its own.