Industry Perspectives

Haiku 5.5 Pricing: What Happens Above 100K Tokens?

The prompt-length boundary changes input, output and cache rates. Inspect a request-level budget, then prototype a worksheet with Atoms.

Start building for free
6 min readPublished
Claude Haiku5.5 title above a ruler marking the100K boundary and two stacks of paper.
On this page

Claude Haiku 5.5 pricing changes when a request's prompt exceeds 100,000 tokens. The higher tier applies to the request, including its output, rather than only the input beyond the boundary. Caching can reduce the bill, but cached input still counts toward prompt length. For a long-running agent, budget from the complete request and its actual cache categories before treating the headline price as your effective rate.

For a prompt described as 100K tokens, verify the exact counted length rather than a rounded label.

Read the two price bands together

The following first-party Claude API rates were checked on October 15, 2026. Prices are USD per million tokens for standard global routing. Partner platforms, paid tools and other modifiers require their own budget.

Billing category Prompt up to 100,000 tokens Prompt over 100,000 tokens
Ordinary input $0.10 $0.50
Output $0.50 $2.50
Five-minute cache write $0.125 $0.625
One-hour cache write $0.20 $1.00
Cache read $0.01 $0.05

Sources: Haiku 5.5 overview and Claude API pricing. The pricing guide specifies that all input tokens, including cache reads and writes, contribute to the threshold. Each request is evaluated separately.

That distinction matters when an agent reads files or receives tool results between turns. A short user message does not make the complete prompt short. Conversely, one long request does not retroactively increase the prices of earlier requests. Attach the price band to each usage row rather than applying a session-wide rate after the fact.

Do not equate the model's context capacity with its cheapest band. Capacity answers whether material fits; pricing answers what processing it costs. A request can be within the supported window and still use the higher rate.

What crossing the boundary costs

Consider synthetic requests with 5,000 output tokens and no cache. The input differs only in length. These are executed arithmetic examples, not measured model runs or promises about the output needed for your task.

Uncached example Input cost Output cost Request total
99,000 input tokens $0.009900 $0.002500 $0.012400
100,000 input tokens $0.010000 $0.002500 $0.012500
100,001 input tokens $0.0500005 $0.012500 $0.0625005
120,000 input tokens $0.060000 $0.012500 $0.072500

Formula: input_tokens × input_rate / 1,000,000 + output_tokens × output_rate / 1,000,000. The rates come from the table above. At the boundary, the illustrative total rises from $0.012500 to $0.0625005. This is a consequence of the tariff applied to the whole request, not a claim that one extra token makes the model do proportionally more work.

For budgeting, identify requests near the threshold and allow room for tool schemas, returned documents and conversation history. Do not discard essential evidence to force a request into the cheaper tier. If trimming changes the answer's reliability, compare the cost of accepted work instead of the initial request price.

A cache hit can still use the higher band

Suppose a request contains an unchanged 110,000-token reference prefix, 10,000 fresh tokens and 5,000 output tokens. Its total input is 120,000 tokens whether the prefix is written to cache or read from cache. It therefore belongs in the higher band.

Cache scenario Calculation Total
No cache 120,000 × 0.50 + 5,000 × 2.50, divided by 1M $0.072500
Initial five-minute cache write 110,000 × 0.625 + 10,000 × 0.50 + 5,000 × 2.50, divided by 1M $0.086250
Later successful cache read 110,000 × 0.05 + 10,000 × 0.50 + 5,000 × 2.50, divided by 1M $0.023000

One write followed by four successful reads totals $0.178250 in this synthetic example; five uncached requests total $0.362500. The difference is $0.184250. It assumes the prefix is eligible, unchanged and reused within the cache lifetime. A missed or expired cache would require different usage categories and a different total.

The prompt-caching documentation explains matching and retention. Use reported cache usage to populate the worksheet. Do not infer a hit from a fast response or from having submitted similar words earlier.

Decide whether to trim, split or keep the context

Start with the task's evidence requirements. For extraction from a specific document section, supply that section and enough surrounding context to preserve meaning. For an analysis that genuinely depends on the entire record, keep the record and budget the higher band.

Splitting is appropriate only when the result can be checked after recombination. For example, independently extracting invoice rows may be separable; resolving references between sections may require shared context. Ask what information a chunk would lose before treating chunking as a cost optimization.

A practical comparison keeps the source version and acceptance criteria fixed:

  1. Run the full-context approach on an authorized sample.
  2. Try a smaller relevant input or a documented caching configuration.
  3. Check the same required fields and source citations.
  4. Include retries and corrections in the cost record.
  5. Keep the cheaper approach only if it still clears the task's quality bar.

This procedure is a proposed measurement method. The arithmetic above does not establish that either approach produces better answers, needs fewer corrections or reaches a particular latency. Those conclusions require your own observed results.

Keep an invoice-compatible request log

Store each request's model, provider, source version, ordinary input, cache-write input, cache-read input, output and applicable price date. Keep unknown usage blank rather than substituting zero. A zero denotes an observed absence; an empty field denotes missing evidence.

Calculate prompt length by adding the input categories once. Then select the band before calculating category costs. Avoid adding cached input to a total that already includes it; that would double-count the prefix. Likewise, keep a provider's billed output quantity separate from the number of visible words in the answer.

For a team, add the task ID and acceptance result. That lets a reviewer distinguish a low-cost failed attempt from a slightly more expensive accepted result. Preserve failed spend instead of deleting unsuccessful rows. A useful monthly forecast reflects the mix of request lengths and repeat reads you actually observe.

Build a reviewable budget workspace

For repeated cost decisions, Atoms' AI app builder can help you prototype a worksheet interface from a natural-language brief. Use synthetic records first and review its calculations independently.

text
Build a Haiku 5.5 request-budget app using synthetic usage rows.
Store ordinary input, five-minute writes, one-hour writes,
cache reads and output separately. Calculate total prompt length
from input categories once. Select the up-to-100,000 or
over-100,000 band for each request. Show every rate and formula.
Keep missing usage unknown. Separate task acceptance from cost.
Allow CSV import/export. Do not connect provider credentials.

Validate the generated app with the examples in this article: exact boundary, one token above it, cache write and cache read. Check that editing a usage category updates both the band and the total. The prompt-caching guide provides a broader measurement method for stable prefixes.

Atoms supports building and iterating websites and web applications; this proposed worksheet does not establish a native Haiku integration or a completed production billing tool. Authentication, data persistence and any live integration need their own review when you move beyond the synthetic prototype.

Conclusion

Budget Haiku 5.5 per complete request. Count cached input toward the 100,000-token threshold, apply the selected rates to the actual billing categories, and include corrections when comparing workflows. Build the worksheet in Atoms if your team needs an editable way to inspect those assumptions before connecting real usage.

A little more clarity

Frequently asked questions

01Q1: Does the higher price apply only to tokens beyond 100K?

No. Claude's pricing guide describes a request-level band. Once the prompt exceeds 100,000 tokens, the request uses the higher input, output and caching rates. Earlier requests retain their own rates.

02Q2: Can caching keep a 120K prompt in the cheaper band?

No. Cache reads and writes count toward prompt length. Caching may reduce category costs while the request remains in the higher band. Use actual reported cache categories to calculate that reduction.

03Q3: Is the output price determined by output length?

For these two bands, prompt length determines which output rate applies. The amount charged still depends on the billed output tokens. A short answer to a long prompt uses the long-prompt output rate.

04Q4: Are these results a model benchmark?

No. They are executed cost calculations with synthetic inputs. They do not measure latency, answer quality or token use for a real task. Keep those measurements in separate columns when testing your application.

Share this article
Made with Atoms

Your next idea starts here.

Turn what you learned into a working app or website.

Start building for free