How to get the most from your Claude startup credits

The program's API credits are $1,000 and they expire six months after they are granted. That is about $165 a month. These are the levers Anthropic documents for stretching them.

1. Plan against the expiry

• Divide the credits by the months you have: $1,000 over six months is roughly $165 a month, or about $5.50 a day.
• Credits work only on the first-party Claude API, not on AWS Bedrock or Google Vertex AI, so build the thing you will actually ship on the Claude API.
• Watch usage in the Claude Console. Check it weekly for the first month to learn your real burn rate.

2. Match the model to the task

Anthropic's own guidance: Haiku for simple tasks, Sonnet for most production workloads, Opus for the most complex reasoning. The price gap is large. Official per-million-token prices on 2026-10-08:

Prototype with a cheaper model, then move only the steps that need it to a larger one. Newer models use a tokenizer that produces about 30% more tokens for the same text than Claude Sonnet 4.6 and earlier, so re-measure cost when you switch.

3. Turn on prompt caching

If many requests share a long prefix (a system prompt, tool definitions, a document), the API can reuse it instead of reprocessing it.

• A cache read costs 0.1x the base input price on most models (0.05x on Claude Opus 5.5 and Sonnet 5.5).
• Writing to the cache costs 1.25x for the 5-minute lifetime, or 2x for the 1-hour lifetime. The 5-minute cache pays off after one read.
• Short prompts are not cached: the minimum is 512 tokens on newer models, higher on older ones.
• Easiest start: add a single top-level cache_control field and let the system place breakpoints.

4. Use the Batch API for work that can wait

The Message Batches API gives a 50% discount on input and output tokens. Most batches finish in under an hour. Good fits: evaluations, bulk classification, back-filling content, nightly jobs. It is not for anything a user is waiting on. Batch and caching discounts can be combined.

5. Watch the extras

• Web search is billed at $10 per 1,000 searches on top of tokens.
• Tool definitions count as input tokens on every request.
• Output tokens cost several times more than input. Cap output length where you can.

A worked example

On Claude Sonnet 5.5 at $2 per million input tokens, $1,000 buys 500 million input tokens at the base price. With cache reads at $0.10 per million, the same budget covers far more cached input. Your real mix of input, output and cached tokens decides the result, so measure it in the Console.

See also credits and perks and how other programs compare.