Haiku 5.5 tiered pricing calculator
The live calculator starts with 100,000 total prompt tokens, 10,000 output tokens and an 80,000-token cacheable prefix. Edit the quantities to update the cost immediately; no Calculate button is required. The worked example below is also included in the initial HTML.
Use total prompt tokens, including cached input. At exactly 100,000 input tokens Haiku 5.5 uses the lower schedule. At 100,001 it uses the higher schedule for all input categories and output, rather than charging only the extra token at a higher rate. The comparison holds token quantities constant across models.
Example result: Haiku 5.5 vs Haiku 4.5 pricing
Example result for one request: 100,000 prompt tokens and 10,000 output tokens; cache scenarios allocate 80,000 input tokens to the cache and 20,000 to ordinary input. Cache writes use the 5-minute duration. The fixed example remains readable without JavaScript; live results use your edits.
Facts checked on 2026-10-09. Each cost below links to the official rate source. This compares equal token quantities, not equal text: actual Haiku 5.5 token counts can differ.
| Scenario | Haiku 5.5 / request | Haiku 4.5 / request | Difference / request | Lower cost vs 4.5 |
|---|---|---|---|---|
| No cache | $0.015 USD · checked 2026-10-09 · Anthropic pricing | $0.15 USD · checked 2026-10-09 · Anthropic pricing | $0.135 USD saved · checked 2026-10-09 · Anthropic pricing | 90% |
| Cache write (5m) | $0.017 USD · checked 2026-10-09 · Anthropic pricing | $0.17 USD · checked 2026-10-09 · Anthropic pricing | $0.153 USD saved · checked 2026-10-09 · Anthropic pricing | 90% |
| Cache read (hit) | $0.0078 USD · checked 2026-10-09 · Anthropic pricing | $0.078 USD · checked 2026-10-09 · Anthropic pricing | $0.0702 USD saved · checked 2026-10-09 · Anthropic pricing | 90% |
Haiku 5.5 pricing tiers and Haiku 4.5 pricing
Facts checked on 2026-10-09. Rates are USD per 1M tokens for standard first-party Claude API usage. Every price links directly to Anthropic's official pricing table. Batch discounts, regional modifiers and separately metered tools are outside this calculator.
The Haiku 4.5 column is the comparison baseline. Select the Haiku 5.5 column using total input length; output tokens do not select the tier. Recheck the official source when using this saved rate snapshot.
How prompt caching changes the cost
Choose no cache, cache write or cache read. The cacheable-prefix field is a subset of total prompt tokens, not extra input. A write charges that prefix once at the selected 5-minute or 1-hour write rate; a hit uses the read rate. The rest of the prompt uses ordinary input pricing. A read assumes an existing matching cache entry and excludes its earlier write cost.
Facts checked on 2026-10-09. Haiku 5.5 requires at least 512 tokens in the cacheable prefix; Haiku 4.5 requires 4,096. Below a model's minimum the calculator falls back to ordinary input charges for that model, following the official caching guide. A cache miss also uses ordinary input charges.
Total cost = uncached input × input rate + cache prefix × write or read rate + output × output rate, divided by 1,000,000. Each input token belongs to one category. With caching, the API input_tokens field alone omits cache reads and writes; add all three input usage fields before selecting the price tier.
| Caching rule | Official documentation |
|---|---|
| Minimum prefix and cache-hit behavior | Anthropic prompt caching · checked 2026-10-09 |
| Input, write and read usage fields | Anthropic usage breakdown · checked 2026-10-09 |
Haiku 4.5 vs Haiku 5.5 tokenizer estimates
Paste the same text into the comparison below. This browser does not contain either official Haiku tokenizer: it uses cl100k_base as a proxy baseline, then rounds that baseline × 1.30 up to a whole token for Haiku 5.5. Both model rows are estimate / reference-only. The raw encoding row is exact for cl100k_base text only.
Facts checked on 2026-10-09. Anthropic's Haiku migration guide describes roughly 30% more tokens for the same text, with content-dependent variation. This illustrative multiplier is not a measured Haiku tokenizer difference and must not determine the 100k billing boundary. Use count_tokens separately for each model, or actual usage, for a model-aware comparison.
| Example view | Tokens | Support / method |
|---|---|---|
| Raw cl100k_base | 37 | exact for this encoding only |
| Haiku 4.5 proxy | 37 | estimate / reference-only: raw reference count |
| Haiku 5.5 illustration | 49 | estimate / reference-only: ceil(reference × 1.30) |
| Illustrated difference | +12 (32.43%) | Rounding changes the displayed percentage; actual model counts can differ |
Summarize this customer request in three bullet points, identify the next action, and return a short JSON response. Customer: I need to change my delivery address before the order ships.Token counts, price tiers and billed usage
- 1. Count the complete supported input separately with claude-haiku-4-5 and claude-haiku-5-5. Include system text, history, supported tools and media in the provider request.
- 2. Enter total input and planned output in this calculator. With caching, total input is input_tokens + cache_creation_input_tokens + cache_read_input_tokens. Do not add cached input twice.
- 3. Reserve output separately: visible answer length does not predict reasoning or final billed output. This calculator does not infer missing chat formatting, images or tool usage from pasted text.
- 4. After the request, reconcile token categories with Messages usage and the official rates. The local text proxy is for exploring assumptions; it is not an invoice or a Claude tokenizer.
Tiktokenizer Reference Sources
Tokenizer behavior and model limits change. Verify production decisions with current provider documentation:
- Anthropic pricing: both Haiku versions and every cache rate · checked 2026-10-09
- Anthropic prompt caching: usage fields and minimum prefixes · checked 2026-10-09
- Haiku 5.5 migration guide: tokenizer change · checked 2026-10-09
- Anthropic token counting: model-aware preflight estimates · checked 2026-10-09