Ninety, Seventy-Five, Twenty-Five: Three Discounts on One Haiku

Anthropic released Claude Haiku 5.5 on 7 October, and the launch post makes its price claim in one sentence: “On average, it now costs around 75% less to run.” A footnote explains the average. Haiku 5.5 “is priced 90% lower than Claude Haiku 4.5 for requests up to 100,000 tokens, and 50% lower for requests over 100,000 tokens. On Haiku 4.5, 90% of requests fell into the former category.” The footnote also says the calculation accounts for a new tokenizer, which means the model “uses slightly more tokens per task.”

So there are three numbers in the room. The rate card says 90%. The headline says 75%. Artificial Analysis, whose model page I read on 7 October, has it at $0.21 per task against $0.28 for Haiku 4.5, which is 25% less. I think all of them are honest. They are also answers to different questions, and the gap between them is a decent map of what a per-token price does and does not tell you.

The rate card is the clean part

Start with the prices, because they are unusually tidy. Anthropic’s pricing documentation lists Haiku 5.5 in two rows. For prompts up to 100,000 tokens it charges $0.10 per million input tokens, $0.50 for output, $0.125 for a five-minute cache write, $0.20 for a one-hour cache write and $0.01 for a cache read. For prompts over 100,000 tokens every one of those numbers is five times higher: $0.50, $2.50, $0.625, $1 and $0.05. Haiku 4.5 has one row: $1, $5, $1.25, $2 and $0.10.

Divide one row by the other and the ratio is exactly 0.1 on every line in the short tier and exactly 0.5 on every line in the long tier. That matters for the arithmetic below. Whatever share of your bill goes to cache reads, to output or to fresh input, the discount within a tier does not move. Only two things can move the total: how much of the bill sits above the 100,000-token line, and how many tokens the new model spends on the same job.

Claude Haiku 4.5 and Haiku 5.5 list prices by line and tier Dot plot on a log scale of prices per million tokens. Input: Haiku 4.5 $1, Haiku 5.5 $0.10 for prompts up to 100,000 tokens and $0.50 above. Output: $5, $0.50 and $2.50. Five-minute cache write: $1.25, $0.125 and $0.625. Cache read: $0.10, $0.01 and $0.05. Every short-tier price is one tenth of Haiku 4.5's and every long-tier price is one half. Source: Anthropic pricing documentation, 7 October 2026. PRICE PER MILLION TOKENS · LOG SCALE · 7 OCT 2026 $0.01 $0.10 $1 $10 Input Output Cache write (5 min) Cache read Input, Haiku 4.5: $1 — Anthropic, 7 Oct 2026 Output, Haiku 4.5: $5 — Anthropic, 7 Oct 2026 Cache write, Haiku 4.5: $1.25 — Anthropic, 7 Oct 2026 Cache read, Haiku 4.5: $0.10 — Anthropic, 7 Oct 2026 Input, Haiku 5.5 over 100k: $0.50 — Anthropic, 7 Oct 2026 Output, Haiku 5.5 over 100k: $2.50 — Anthropic, 7 Oct 2026 Cache write, Haiku 5.5 over 100k: $0.625 — Anthropic, 7 Oct 2026 Cache read, Haiku 5.5 over 100k: $0.05 — Anthropic, 7 Oct 2026 Input, Haiku 5.5 up to 100k: $0.10 — Anthropic, 7 Oct 2026 Output, Haiku 5.5 up to 100k: $0.50 — Anthropic, 7 Oct 2026 Cache write, Haiku 5.5 up to 100k: $0.125 — Anthropic, 7 Oct 2026 Cache read, Haiku 5.5 up to 100k: $0.01 — Anthropic, 7 Oct 2026 $0.10 $0.50 $1 Haiku 5.5, prompt ≤100k (÷10) Haiku 5.5, prompt >100k (÷2) Haiku 4.5
On a log scale every row has the same shape. The short-tier price sits one decade below Haiku 4.5, and the long-tier price sits a factor of two below it. Cache reads get no special treatment.

What the 75% has to assume

The footnote’s 90% is a share of requests, but a bill is weighted by tokens, and a request over 100,000 tokens is by definition bigger than one under it. Ten percent of requests can easily be a quarter or a third of the money. Call \(s\) the share of an old Haiku 4.5 bill that came from long prompts, and call \(m\) how many more tokens Haiku 5.5 uses for the same work. Then the new bill, as a fraction of the old one, is \((0.1 + 0.4s) \times m\). That is my framing, not Anthropic’s, but it falls straight out of the rate card.

Plug in the footnote’s numbers literally (\(s = 10\%\), \(m = 1\)) and the saving is 86%. Anthropic’s migration guide puts a figure on the tokenizer: “the same input text produces approximately 30% more tokens on Claude Haiku 5.5 than on Claude Haiku 4.5,” with the exact increase depending on content. With \(m = 1.3\) and \(s\) still at 10%, the saving is about 82%. To get down to 75%, one of two things has to be true. Either long prompts made up about 23% of the old bill, which is entirely plausible for 10% of requests, or the model spends about 79% more tokens per task, which is a long way from “slightly more.” Anthropic doesn’t publish either input, so I can’t say which, and the real answer is probably some of each, plus output lengths that move with the new effort setting.

Haiku 5.5 saving versus Haiku 4.5, by long-prompt share of the bill Two lines show the cost reduction as the share of the old bill from prompts over 100,000 tokens rises from 0% to 60%. With the same token count the saving falls from 90% to 66%; with 30% more tokens it falls from 87% to about 56%. At a 10% long-prompt share the savings are 86% and about 82%. Anthropic's headline of 75% is reached at a 37.5% share with no token change, and at about 23% with 30% more tokens. Writer's arithmetic from Anthropic's published prices and tokenizer note, 7 October 2026. COST REDUCTION VS HAIKU 4.5 · BY LONG-PROMPT SHARE OF THE OLD BILL 90% 80% 70% 60% 50% 0% 10% 20% 30% 40% 50% 60% Anthropic's "around 75% less" Footnote's 10%, read as share of spend Same tokens, 10% long share: 86% less — writer's arithmetic 30% more tokens, 10% long share: about 82% less — writer's arithmetic Same tokens: 75% reached at a 37.5% long share — writer's arithmetic 30% more tokens: 75% reached at about a 23% long share — writer's arithmetic 86% 82% 37.5% 23% same tokens 30% more tokens Writer's arithmetic: new bill ÷ old = (0.1 + 0.4 × long share) × token multiplier. Share is of spend, not requests.
Reading the footnote's 10% as a share of spend gives 86%, or about 82% once the tokenizer is counted. Getting to 75% needs long prompts to carry roughly a quarter of the old bill, or a much bigger token increase than the footnote's "slightly more."

There is a second-order effect hiding in the tokenizer. The tier boundary is counted in Haiku 5.5 tokens. If the same text now produces about 30% more tokens, the 100,000-token line falls at roughly 77,000 of the old tokens. A request that sat comfortably in the cheap bucket on Haiku 4.5 at 85,000 tokens lands in the expensive one on Haiku 5.5. The footnote’s 90% describes Haiku 4.5 traffic and doesn’t say which tokenizer it was counted in. If it was the old one, the share of requests over the line on the new model will be somewhat higher than 10%. How much higher depends on how much traffic sits between 77k and 100k, which only Anthropic knows.

Nor does the documentation say what counts as “the prompt.” The pricing page says only that “a prompt of over 100,000 tokens pays higher prices.” For a chat request that is unambiguous. For an agent whose context is mostly a cache read of everything it has done so far, it matters a great deal whether those cached tokens count toward the line. If they do, a long-running agent crosses 100,000 tokens early and stays there. Whether 100,000 tokens is low enough to matter for agents was one of the arguments in the Hacker News thread. One commenter, minimaxir, called 100k “an absurdly low cutoff” that “will be quickly exceeded if you are doing anything with Agents.” Another commenter replied by quoting the 90% footnote. Both are right, about different workloads.

The surcharge is also new for this model. The same pricing page says that every Claude model from 4.6 onward “(except Claude Haiku 5.5)” bills its full million-token context at one rate. Anthropic has split prices by prompt length before. When Sonnet 4 got a million-token context window in August 2025, prompts over 200,000 tokens paid twice the input rate and 1.5 times the output rate. Haiku 5.5’s step is a factor of five on every line, and it starts at half the length.

The 25% is a different question

Artificial Analysis publishes a cost per task for each model it runs through its Intelligence Index. On 7 October it listed Haiku 5.5 at maximum effort, in an entry it labels “Max, Default Fallback,” at $0.21 a task and Haiku 4.5 with reasoning at $0.28. The reason the drop is a quarter and not three quarters is in the next line of each page. To run the index, Haiku 5.5 generated 440 million output tokens, which Artificial Analysis calls “very verbose in comparison to the median of 100M.” Haiku 4.5 generated 78 million. That is about 5.6 times as many output tokens on the same set of tasks.

It also scored 43 on the index against Haiku 4.5’s 17, so this isn’t waste in any simple sense. The model is doing more work per task and getting more of them right. Anthropic’s own table tells the same story more loudly: Haiku 4.5 scored 0.0% on Terminal-Bench 4.0 and Haiku 5.5 scored 39.2%. Going from 0.0% to 39.2% on Anthropic’s own run is less a cost comparison than a category change. Still, it means the 75% figure and the 25% figure are measuring different things. Anthropic’s number holds the work roughly constant and asks what the same traffic costs. Artificial Analysis lets the model think as long as its top effort setting allows and asks what a task costs. Haiku 5.5 is Anthropic’s first Haiku with an adjustable effort setting, so some of the gap between those two numbers is presumably a dial the customer now controls. Neither source says how much.

Haiku 4.5 and Haiku 5.5 on the Artificial Analysis Intelligence Index Three slope panels from Artificial Analysis, 7 October 2026, each comparing Haiku 4.5 with reasoning against Haiku 5.5 at maximum effort. Output tokens to run the index rise from 78 million to 440 million. Cost per task falls from $0.28 to $0.21. Index score rises from 17 to 43. ARTIFICIAL ANALYSIS INTELLIGENCE INDEX · 7 OCT 2026 Output tokens Cost per task Index score Haiku 4.5 (reasoning): 78M output tokens — Artificial Analysis, 7 Oct 2026 Haiku 5.5 (max): 440M output tokens — Artificial Analysis, 7 Oct 2026 Haiku 4.5 (reasoning): $0.28 per task — Artificial Analysis, 7 Oct 2026 Haiku 5.5 (max): $0.21 per task — Artificial Analysis, 7 Oct 2026 Haiku 4.5 (reasoning): score 17 — Artificial Analysis, 7 Oct 2026 Haiku 5.5 (max): score 43 — Artificial Analysis, 7 Oct 2026 78M 440M $0.28 $0.21 17 43 Haiku 4.5 Haiku 5.5 Haiku 4.5 Haiku 5.5 Haiku 4.5 Haiku 5.5 Each panel on its own linear scale from zero. Haiku 4.5 with reasoning; Haiku 5.5 at maximum effort.
At its top effort setting Haiku 5.5 wrote about 5.6 times as many output tokens as Haiku 4.5 on the same index. That ate most of a tenfold price cut and left the cost per task 25% lower, for a score two and a half times higher.

Averages over mixtures

A blended figure is the only honest way to state one number for a two-tier price, and Anthropic printed its weights, which is more than launch posts usually do. The trouble with blended figures is older than language models. In 1975, using the autumn 1973 admissions, Peter Bickel, Eugene Hammel and J. William O’Connell published a study of graduate admissions at Berkeley in Science. Pooled across the university, the numbers looked biased against women. Department by department, the bias mostly disappeared, because women had applied in larger numbers to the departments that rejected most applicants. The aggregate was accurate. It just answered a question about the mix, not about any department.

A per-token discount works the same way. Haiku 5.5 is 90% cheaper for a short classification call. It is 50% cheaper for a long document summary. It is somewhere else entirely for an agent that reads its own growing context back on every turn and is set to think hard. The 75% is the average over Anthropic’s view of Haiku 4.5’s traffic, and the only bill it describes exactly is that one.

The same post halves Sonnet 5.5’s cache-read price, from $0.20 to $0.10 per million tokens, and says that makes it “around 20% cheaper on most agentic work.” By the same arithmetic I used on Opus 5.5 and GPT-6 Sol, halving one line saves 20% only if that line was about 40% of the bill. That is a reasonable share for an agent and a high one for a chatbot. The pattern is the one I wrote about with Jev’s speed claims: the number on the page is true, and the denominator is the part you have to bring yourself.

If I were pricing a move to Haiku 5.5, I would skip both headlines and pull two numbers from my own logs. The first is the share of Haiku 4.5 spend that came from prompts above about 77,000 old-tokenizer tokens. The second is output tokens per task at the effort level I would actually run. The rate card does the rest.

References

  1. Anthropic (2026). Claude Haiku 5.5. Launch post, 7 October 2026.
  2. Anthropic (2026). Pricing. Claude Platform documentation, accessed 7 October 2026.
  3. Anthropic (2026). Migrating to Claude Haiku 5.5. Claude Platform documentation, accessed 7 October 2026.
  4. Anthropic (2025). Claude Sonnet 4 now supports 1M tokens of context. 12 August 2025.
  5. Artificial Analysis (2026). Claude Haiku 5.5: Intelligence, Performance & Price Analysis. Accessed 7 October 2026.
  6. Artificial Analysis (2026). Claude 4.5 Haiku (Reasoning): Intelligence, Performance & Price Analysis. Accessed 7 October 2026.
  7. Hacker News (2026). Claude Haiku 5.5. Discussion thread, 7 October 2026; comment by minimaxir.
  8. Bickel, P. J., Hammel, E. A. & O’Connell, J. W. (1975). Sex Bias in Graduate Admissions: Data from Berkeley. Science 187(4175): 398–404, 7 February 1975.