Official pricing source notes: Claude Haiku 5.5

Source: https://platform.claude.com/docs/en/about-claude/pricing
Cross-check: https://www.anthropic.com/claude-haiku-5-5
Retrieved: 2026-10-08

Provenance: These are transcribed factual pricing fields and paraphrased rules from the official Anthropic pages, retrieved through the external web reader. They are not a raw webpage download, API response, invoice, or measured usage. The Open-Science session could not retrieve the page directly because its network address check rejected the resolved address. This attachment preserves the inputs for the calculation without changing that network check.

Scope: Anthropic first-party standard API pricing, USD per 1,000,000 tokens. Other providers, taxes, negotiated discounts, server-tool charges and regional modifiers are not included.

Prompt length tier,Input,Output,5-minute cache write,1-hour cache write,Cache read
Up to and including 100000 tokens,0.10,0.50,0.125,0.20,0.01
Over 100000 tokens,0.50,2.50,0.625,1.00,0.05

The official model pricing table lists these as separate prompt-length tiers. The long-context pricing section states that Haiku 5.5 has higher rates for prompts longer than 100000 tokens. The published structure is a request-length tier, rather than a table giving an incremental surcharge only on excess tokens.

Prompt caching rules from the official pricing page:
- Creating a five-minute cache is priced at 1.25 times the base input rate; a one-hour cache is priced at twice that rate.
- Reading cached input uses the cache-read price. Storing it on the first request uses the cache-write price; do not count a first write as a hit.
- A hit/refresh retains the duration corresponding to the earlier cache write.
- Batch processing has separately published discounted rates. Keep standard synchronous and batch scenarios separate.

Calculation boundaries:
- This attachment establishes the published rate table, not the number of tokens in any actual document.
- Exact request composition, cache eligibility, hit behavior and expiry must be stated as assumptions when constructing examples. A cache hit is not guaranteed merely because a request repeats.
- Do not silently invent additional API metering rules that are not established by the sources. If a cache-related threshold detail remains unverified, label that scenario conditional and retain the uncached calculation as the directly supported core example.
- All calculated dollar amounts are estimates from published rates, not bills observed from a paid model call.
