Claude Haiku 5.5 Is 90% Cheaper on Paper. What Will Your Bill Do?

Anthropic released Claude Haiku 5.5 on 7 October 2026 with a headline that any budget owner will notice: input tokens at $0.10 per million, down from $1 on Haiku 4.5. So the question is simple. If you run Haiku 4.5 today, will switching cut your bill by 90%?

The short answer

Probably not by 90%, and not by the same amount for every workload. The cut is real for short prompts, but three details in Anthropic's own documentation pull the average down: a tokenizer that counts the same text as more tokens, a price that roughly quintuples once a prompt passes 100,000 tokens, and thinking that is switched on by default and billed as output. Anthropic itself puts the average saving at about 75%, not 90%. Your number depends on prompt length, output length and effort setting, so the sensible move is to measure it on your own traffic before you migrate.

What Anthropic announced

According to the launch announcement (7 October 2026), Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Above that threshold the rates are $0.50 and $2.50. Haiku 4.5 is $1 and $5. The pricing documentation adds cache reads at $0.01 per million tokens for the short tier and a 50% batch discount.

The same page makes the average claim explicit: Haiku 5.5 is “around 75% less to run.” A footnote explains the basis. It is 90% cheaper for requests up to 100,000 tokens, 50% cheaper above that, and Anthropic says about 90% of Haiku 4.5 requests fell in the first group. The calculation also accounts for the new tokenizer. The announcement does not show the full arithmetic, so treat 75% as the company's estimate for its own traffic mix, not a guarantee.

Where the 90% shrinks

The tokenizer

The “What's new” documentation says Haiku 5.5 uses the same tokenizer as Claude 4.7 and later models, so the same text counts as approximately 30% more tokens than on Haiku 4.5, with the exact increase depending on content. Anthropic's advice is to recount prompts and revisit cost estimates. A price per token is only half of a price per task.

The 100,000-token step

Long context is no longer flat-priced on this model. The pricing page says a request whose prompt exceeds 100,000 tokens pays the higher rates, and that the prompt length counts all input tokens, including cache reads and cache writes. Each request is priced on its own, so a cache hit does not protect a request that crosses the line.

Combine that with the tokenizer and a quirk appears. By my arithmetic, using Anthropic's “approximately 30%” figure, a document that measured about 77,000 tokens on Haiku 4.5 would measure about 100,000 on Haiku 5.5. Anyone who sat comfortably under the threshold on the old model could land on the wrong side of it on the new one. A 100,000-token Haiku 4.5 prompt would become roughly 130,000 tokens, billed at $0.50 per million: about $0.065 against $0.10 before, a saving of around a third rather than 90%. This is an illustration, not a measurement; real ratios vary.

Thinking is billed as output

Adaptive thinking is on by default, and the docs state that thinking counts toward output tokens even when the thinking text is omitted from the response. Haiku 5.5 also keeps thinking blocks from all earlier turns in context, which the documentation says means multi-turn conversations carry more input tokens than the tokenizer change alone explains.

The effort documentation sets the default at medium and describes five levels, from low to max. Anthropic recommends starting at medium, says low is for simple high-volume requests (with a warning that agents are more likely to skip a search or a check at low), and reserves xhigh and max for cases where your own evaluations show a gain. Effort is the main dial on how many output tokens you pay for.

A worked example

Take a classification call with a 10,000-token prompt on Haiku 4.5 and a 500-token reply. At $1 and $5 per million, that is $0.0125. On Haiku 5.5 the same text becomes about 13,000 input tokens ($0.0013) and, say, 650 reply tokens ($0.0003), roughly $0.0016, or close to 87% cheaper. The saving survives because the output is short. Now let the model think for 20,000 tokens first and the output bill alone reaches about $0.0103, which takes the total to roughly $0.0116. At that point you have saved less than 10%. The numbers are mine, built from published list prices, and the thinking volume is hypothetical; the point is that output volume, not the headline input rate, decides whether you land near 90% or near zero.

What the capability claims do and do not show

The announcement includes a benchmark table comparing Haiku 5.5 with Haiku 4.5, Sonnet 5.5 and OpenAI's GPT-6 Luna. Examples: 72.4% on the offline subset of OSWorld 2.1, against 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna; 39.2% on Terminal-Bench 4.0, against 0.0% for Haiku 4.5. These are Anthropic's reported results. The page credits Artificial Analysis only for the GDPval-AA v2.1 score, and sends readers to the system card for other methods. Customer quotes from Asana, HubSpot, Box and others are described as early testing. None of this is an independent replication.

Anthropic is also candid about scope: it says Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding, and positions Haiku 5.5 for narrow tasks such as summarisation, compaction and subagent work. The model page describes it as built for classification, routing, extraction and subagent tasks.

What we know vs. what we're inferring

• Known (Anthropic documentation, 7 October 2026): the list prices, the 100,000-token threshold, the roughly 30% tokenizer increase, default medium effort, thinking billed as output, and the 75% average-saving estimate with its stated basis.
• Anthropic-reported, not independently verified: benchmark scores and customer results. Only one score is attributed to a third party.
• Inferred by us: the 77,000-token crossover and the worked example. They follow from published numbers but are not measurements. We did not run Haiku 5.5 ourselves, and we did not find independent token-consumption data we could verify at the time of writing.

What to watch next

• Independent cost-per-task data. Evaluators such as Artificial Analysis may publish how many output tokens Haiku 5.5 spends at each effort level. That will say more about real bills than any price table.
• Your own replay test. Re-run a sample of production prompts, count tokens with the model's own token counter, and compare cost per completed task at low, medium and high effort.
• Prompt-length distribution. Check how many of your requests sit between roughly 75,000 and 100,000 tokens on Haiku 4.5. Those are the ones the threshold may catch.
• Competitor responses. Anthropic's own table treats GPT-6 Luna as the reference rival; pricing moves from OpenAI or Google would change the comparison.
• Retirement timing. Anthropic lists Haiku 5.5 as not retiring before 7 October 2027, so there is time to test rather than rush.

Educational information only.