What changes at 100,000 prompt tokens?
Haiku 5.5 has a per-request pricing boundary, not a marginal surcharge. A prompt containing exactly 100,000 tokens stays in the lower tier. A prompt containing 100,001 tokens enters the higher tier for both input and output.
That means adding one prompt token can change the price of the whole request. It does not reprice earlier requests, and monthly token volume does not decide which tier an individual request uses.
The chart and boundary examples use identical output lengths to isolate that effect. They are calculated estimates from published rates, not observed charges, invoices, latency results or quality measurements.
Scope: direct Claude API, standard global synchronous requests, before taxes and negotiated discounts. Batch processing, regional routing and third-party hosting are outside these examples. See the Anthropic pricing page for the broader model lineup.
Caching saves money, but does not remove the boundary
Anthropic counts uncached input, cache reads and cache writes toward the request’s prompt length. Count these categories together when deciding the tier, then price each token category at that tier’s applicable rate.
For example, a request containing 90,000 cached tokens and 11,000 uncached tokens has a 101,000-token prompt. Its cache reads do not make it a short-prompt request. Output tokens are charged at the selected output rate; they do not count toward this prompt-length boundary.
Cache reads, five-minute writes and one-hour writes have distinct rates. A cold cache write is not a cache hit. Record actual usage categories, and never assume every repeated prefix receives a read discount.
For broader guidance, read cached tokens and AI cost savings.
What to do in a production app
- Budget per request. Use Anthropic’s official token-counting tools for the selected model before a boundary-sensitive call. Character counts and another model’s tokenizer are not substitutes.
- Leave headroom. Include system instructions, conversation history and other prompt content in your budget. A growing conversation can cross the boundary without a longer final user message.
- Track cache categories. Log uncached input, cache-creation input, cache-read input and output separately. Reconcile their definitions with the provider’s usage documentation before estimating costs.
- Trim deliberately. Remove irrelevant retrieval results or old history where task quality permits. Splitting a task is not automatically equivalent: compare answer quality before assuming chunking is a saving.
- Set separate cost and quality checks. This documented rate cliff says nothing about a sudden quality or latency change at the same boundary.
Use our token calculator for planning, with the documented request-length tier selected by the data. A monthly total alone cannot determine a context tier. If you compare alternative providers, check OpenAI pricing separately rather than assuming it shares Anthropic’s threshold.
Is this an increase over earlier Haiku models?
Do not treat this as a blanket increase over earlier models. The sharp increase explained here is between Haiku 5.5’s own two prompt-length tiers, not a before-and-after price change.
This article does not establish when the tier first appeared or that an older article was inaccurate when published. For the wider comparison, see our Claude API pricing guide.
For manual chat rather than an API integration, you can also compare ChatLLM as a multi-model workspace. It is not a substitute for this API rate card; check its current models and limits. Compare manual-chat access. This article contains affiliate links; we may earn a commission. Anthropic API pricing is independent of this referral.
Sources and verification limits
The official pricing documentation and Haiku 5.5 overview agree on the prompt-length tiers. The generated rate table displays DATA’s source-verification timestamp. Daily source checks watch for pricing changes and stale records; this is not a guarantee of real-time prices.
We have not tested Anthropic’s billing empirically. Demonstrating actual charges would require authorized API runs and reconciliation with attributable billing records. Usage multiplied by a rate card remains an estimate, not independent billing proof.