Claude Haiku 5.5 Charges Five Times More Above 100,000 Tokens: What One Request Costs on Each Side of the Line

Claude Haiku 5.5, which Anthropic launched on October 7, 2026, is the only model on Anthropic’s pricing page whose price depends on how long the prompt is. Up to 100,000 tokens it is the cheapest model Anthropic sells. One token past that, every price on the request is five times higher. This post reads Anthropic’s pricing page, the Haiku 5.5 model page and its migration guide as they stood on October 11, 2026, and works out what the threshold does to the cost of one request, a million requests and a long prompt split in two. It compares prices only: nothing here measures how well the model answers, and this site does not measure that either.

The two price lists, as Anthropic prints them

Anthropic’s pricing page lists Claude Haiku 5.5 on two lines of its “Model pricing” table:

• $0.10 / MTok for prompts up to 100,000 tokens | $0.50 / MTok | $0.125 / MTok | $0.20 / MTok | $0.01 / MTok |
• $0.50 / MTok for prompts over 100,000 tokens | $2.50 / MTok | $0.625 / MTok | $1 / MTok | $0.05 / MTok |

The columns are input, output, 5-minute cache write, 1-hour cache write and cache read. The model page prints the same ten figures, one line per price (“Input $0.10 / MTok for prompts up to 100,000 tokens $0.50 / MTok for prompts over 100,000 tokens”, and so on), and summarises them in its comparison table as “From $0.10 / $0.50”.

Price per 1M tokens	Prompt up to 100,000	Prompt over 100,000	Ratio
Input	$0.10	$0.50	5×
Output	$0.50	$2.50	5×
Cache write, 5 minutes	$0.125	$0.625	5×
Cache write, 1 hour	$0.20	$1.00	5×
Cache read	$0.01	$0.05	5×
Batch input / output	$0.05 / $0.25	$0.25 / $1.25	5×

The ratio is the same on every line, and the output price moves too, even though output tokens play no part in deciding which list applies. The Batch figures come from the page’s Batch table, which prints “$0.05 / MTok for prompts up to 100,000 tokens | $0.25 / MTok” and “$0.25 / MTok for prompts over 100,000 tokens | $1.25 / MTok”, half of each standard figure.

What counts toward the 100,000, and what the price applies to

The pricing page’s “Long context pricing” section has the rule in four sentences: “Claude Haiku 5.5 is priced by prompt length: a request whose prompt is over 100,000 tokens pays higher prices. A request’s prompt length counts all of its input tokens, including cache reads and cache writes. Each request is priced on its own: a request over the threshold pays the higher prices even when part of its prompt is a cache hit, and earlier requests keep the prices they were billed at.”

Three things follow from that wording. A prompt of exactly 100,000 tokens is still on the lower list, because the line reads “up to 100,000”. Cached tokens count toward the length, so caching does not move a request back under the line. And each request stands alone: a conversation that crosses the line on its twentieth turn pays the higher prices from that turn on, not retroactively.

Whether the higher prices apply to the whole request or only to the tokens above 100,000 is the question that decides the arithmetic. The page attaches the prices to the request (“a request whose prompt is over 100,000 tokens pays higher prices”), and the price lines are headed “for prompts over 100,000 tokens”, not as a price for the tokens above 100,000. Nowhere does it describe billing only the excess. It does not use the words “full request” either, which is how OpenAI phrases the same rule for its own models (“for the full request”, on the GPT-6 Luna model page). The figures below follow Anthropic’s wording: a request over the line pays the higher prices on all of its tokens.

One request at 90,000 tokens and one at 110,000

Take two requests that differ only in prompt length, each producing 1,000 output tokens:

• 90,000 input tokens: 90,000 × $0.10 per million is $0.009, plus 1,000 × $0.50 per million, $0.0005. Total $0.0095.
• 110,000 input tokens: 110,000 × $0.50 per million is $0.055, plus 1,000 × $2.50 per million, $0.0025. Total $0.0575.

The second request carries 22% more prompt and costs about 6.05 times as much. Put another way, the $0.055 of input on the 110,000-token request would buy 550,000 input tokens on the lower list. Nothing about the model’s context window stops a request from crossing: Haiku 5.5 has “Context window 1M tokens”, so the line is a price boundary, not a limit.

Cache and Batch on each side

Caching is the obvious way to cut the cost of a long prompt that repeats. On Haiku 5.5 it still works above the line, but from a higher base. Two requests, each with 10,000 fresh tokens and 1,000 output tokens:

Request	Cached part	Fresh input	Output	Total
90,000 tokens, 80,000 read from cache	$0.0008	$0.001	$0.0005	$0.0023
110,000 tokens, 100,000 read from cache	$0.005	$0.005	$0.0025	$0.0125

With most of the prompt cached, the long request costs about 5.4 times the short one rather than 6.05. The cache read is cheap on both sides, so the output and the fresh input, both five times dearer above the line, carry most of the gap. The pricing page says the cache multipliers “stack with other pricing modifiers, including the Batch API discount and data residency”, and the caching page repeats that on a model priced by prompt length, “cache reads and cache writes count toward a request’s prompt length”.

Batch halves both lists and keeps the ratio: the 90,000-token request costs $0.00475 through the Batch API and the 110,000-token request $0.02875. A long-prompt Batch request still costs more than three standard short ones.

Per request and per million requests, at three prompt sizes

With 1,000 output tokens per request, standard and Batch prices from the pricing page read on October 11, 2026:

Prompt	Per request	Per 1M requests	Per request, Batch	Per 1M requests, Batch
20,000 tokens	$0.0025	$2,500.00	$0.00125	$1,250.00
90,000 tokens	$0.0095	$9,500.00	$0.00475	$4,750.00
150,000 tokens	$0.0775	$77,500.00	$0.03875	$38,750.00

From 20,000 to 90,000 tokens the prompt grows 4.5 times and the cost 3.8 times, because output does not grow with it. From 90,000 to 150,000 the prompt grows by two thirds and the cost about 8.2 times. On a million requests, the step from 90,000 to 150,000 tokens adds $68,000.00; at the lower list, with no threshold, the same step would add $6,000.00.

Compared with Claude Haiku 4.5 and other cheap long-context models

Claude Haiku 4.5, now “Active (legacy)” on its model page, has one price list: $1 input and $5 output per million, 200,000 tokens of context. Two cheaper models from other makers in this site’s dataset also price by prompt length: Qwen3.7-Flash, whose Singapore page prints three bands (“Input<=32k”, “32k<Input<=256k”, “256k<Input<=1m”), and GPT-6 Luna, which OpenAI prices higher above 272,000 input tokens “for the full request”. Gemini 2.5 Flash-Lite has a single list and is included as the flat reference. Per request, 1,000 output tokens:

Model	20,000 tokens	90,000 tokens	150,000 tokens	Long-prompt rule	Verified
Claude Haiku 5.5	$0.0025	$0.0095	$0.0775	5× above 100,000	October 11, 2026
Claude Haiku 4.5	$0.025	$0.095	$0.155	none	October 11, 2026
Qwen3.7-Flash	$0.00073	$0.0094	$0.0154	bands at 32k and 256k	October 11, 2026
GPT-6 Luna	$0.0025	$0.0095	$0.0155	2× input, 1.5× output above 272,000	October 5, 2026
Gemini 2.5 Flash-Lite	$0.0024	$0.0094	$0.0154	none	October 5, 2026

Up to 100,000 tokens Haiku 5.5 costs exactly what GPT-6 Luna costs, $0.10 and $0.50, and at 90,000 tokens it is within a hundredth of a cent of the other two. At 150,000 tokens it costs about five times as much as each of them, because their thresholds sit higher (256,000 for Qwen3.7-Flash’s top band, 272,000 for GPT-6 Luna) or do not exist. Qwen3.7-Flash is the cheapest of the five at 20,000 tokens because its first band, up to 32k, costs $0.03 and $0.13. Its prices are Alibaba’s Singapore (“Scope: International”) list.

Haiku 4.5 is the more useful comparison for anyone already on Anthropic. At equal token counts Haiku 5.5 is ten times cheaper below the line and still half the price at 150,000 tokens. But equal token counts are not equal text. The Haiku 5.5 migration guide says “the same input text produces approximately 30% more tokens on Claude Haiku 5.5 than on Claude Haiku 4.5. The exact increase depends on the content.” At exactly 30%, a prompt that measured 90,000 tokens on Haiku 4.5 becomes 117,000 on Haiku 5.5, above the line: with 1,300 output tokens it costs about $0.0618, against $0.095 on Haiku 4.5. Still cheaper, but by about 35%, not 90%. Measured in Haiku 4.5 tokens, the line sits at about 77,000.

A practical rule for splitting long prompts

The arithmetic gives a rule that needs no quality claim. If a job can be split into requests that each stay at or under 100,000 tokens, input and output are both billed at the lower list, one fifth of the higher. Splitting usually repeats something in every request: instructions, a schema, shared context. On input alone, the split stays cheaper as long as all the pieces together, repeats included, add up to less than five times the original prompt. A 150,000-token prompt cut into two pieces of 75,000 tokens, each with a 5,000-token preamble, sends 160,000 tokens instead of 150,000 and costs $0.017 with 1,000 output tokens per piece, against $0.0775 in one request.

Four details decide whether a piece really lands under the line:

• Count with the model’s tokenizer. Use Anthropic’s token counting on Haiku 5.5 itself, not counts from Haiku 4.5 or another model, for the 30% reason above.
• Count everything sent. The rule counts “all of its input tokens, including cache reads and cache writes”, which includes the system prompt, tool definitions and earlier turns. The pricing page’s tool table lists 286 extra system-prompt tokens for Haiku 5.5 when tools are used with auto or none, and 406 with any or tool.
• Watch long conversations. The migration guide says: “Thinking blocks from earlier assistant turns stay in context and count as input tokens, where Claude Haiku 4.5 kept only the latest turn’s.” A conversation grows faster than its visible text.
• Leave a margin. A prompt that sits a few hundred tokens under the line in one request can cross it in the next.

Whether a split job produces the same result as one long request is a question about the model, not the price list, and this post does not answer it. What it can say is the size of the stake: on a million requests a month, the difference between 90,000 and 150,000 tokens is $68,000.00.

How this was read

Every Anthropic figure and sentence comes from the pricing page, the caching page, the Claude Haiku 5.5 model page and its migration guide, downloaded on October 11, 2026, and matches this site’s dataset row for Haiku 5.5, verified the same day. The other models’ prices come from this site’s dataset, where each row links to the page it was read on, with the date shown in the table; GPT-6 Luna’s figures and quote are from OpenAI’s pricing page and model page as recorded there. Prices are standard list prices in US dollars, without data residency premiums. The pricing table has every other model, and the long-prompt ranking compares one long request across all of them.

Frequently asked questions

How much more does Claude Haiku 5.5 charge above 100,000 tokens?

Five times more on every line: $0.50 instead of $0.10 per million input tokens and $2.50 instead of $0.50 per million output tokens, with cache reads going from $0.01 to $0.05, as Anthropic's pricing page read on October 11, 2026 prints them.

Does the higher price apply to the whole request or only to the tokens above 100,000?

Anthropic's page says that a request whose prompt is over 100,000 tokens pays higher prices and that each request is priced on its own; it does not describe charging only the tokens above the line. Read that way, a 110,000-token request with 1,000 output tokens costs $0.0575, against $0.0095 for one of 90,000.

Do cache hits keep a long prompt under the threshold?

No: Anthropic counts cache reads and cache writes in the prompt length, and a request over the threshold pays the higher prices even when part of its prompt is a cache hit. A 110,000-token prompt with 100,000 tokens read from cache still costs $0.0125 with 1,000 output tokens.

Is Claude Haiku 5.5 still cheaper than Claude Haiku 4.5 for long prompts?

At the same token counts, yes: a 150,000-token request with 1,000 output tokens costs $0.0775 on Haiku 5.5 and $0.155 on Haiku 4.5. Anthropic says the same text produces approximately 30% more tokens on Haiku 5.5, which narrows but does not close that gap in the example worked here.