How to cut your Anthropic (Claude) API costs

The example here is 6,600 requests a month (300 a working day), about 15,000 tokens in and 1,000 out each. On Sonnet 5.5 at list price that is $264 a month. Opus costs twice as much per token as Sonnet, and Haiku 5.5 costs a twentieth. Use Opus for the hardest work, Sonnet for most of it, and Haiku for short, simple requests.

With Claude you mark what to cache. A 5 minute cache write costs 1.25 times the input price, and each read after that costs a twentieth of the input price on Opus and Sonnet 5.5, and a tenth on Haiku. So caching pays as soon as the same start is read again before it expires.

Coding tools resend the same files and instructions on every turn, so a large share of input can come from the cache:

Cached rows count cached tokens at the read price and leave out cache writes, so treat them as a floor.

The Message Batches API runs jobs within 24 hours at half the standard price for input and output. It fits evals, backfills, and nightly summaries.

Haiku 5.5 charges more once a prompt passes 100K tokens, so very long prompts can cost more on Haiku 5.5 than you expect. Anthropic also notes that Claude 4.7 and later count more tokens for the same text than older models, so compare the cost of real tasks when you switch.

Set spend limits in the Claude Console, and give each app or person its own key. For ideas that work on every provider, read how to reduce your LLM API costs, or compare Claude with GPT in GPT vs Claude API pricing.