Claude Haiku 5.5 Is Cheap Until 100,000 Tokens. Then 5x.

Anthropic released Claude Haiku 5.5 on 7 October 2026 as “the cheapest, fastest, and most capable small model we’ve ever released”. Input falls from $1.00 to $0.10 per million tokens, output from $5.00 to $0.50 — a 90% cut on the sticker. Anthropic’s own sentence is more careful: “On average, it now costs around 75% less to run.”

The gap is in the pricing table, not the headline. Haiku 5.5 is the only current Claude model priced by prompt length: $0.10 in / $0.50 out up to a 100,000-token prompt, five times that above it. For the loops Haiku is sold for, the question is how many of your requests have already crossed it.

The five rows that multiply by five

From the specifications table and the pricing page, per million tokens:

Batch is 50% off in both columns; the 1.1× inference_geo multiplier applies to the higher one. The same launch halved Sonnet 5.5’s cache reads, $0.20 to $0.10 — Anthropic says that makes Sonnet 5.5 about 20% cheaper on most agentic work.

Anthropic says prompts up to 100,000 tokens “make up around 90% of requests to our previous Haiku model” — traffic to Haiku 4.5, whose window was 200,000 tokens. That is the old model’s traffic, not a guarantee about yours.

A cache hit does not keep you under the line

The pricing page says Claude 4.6 and later models “(except Claude Haiku 5.5) … include the full 1M token context window at standard pricing”. Haiku 5.5 is the exception:

“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.”

The trick that makes long agent loops affordable elsewhere — cache a large stable block, send it every turn — does not pull a request back under the line. Cached tokens count, each turn is priced separately, and a grown context pays the high column on every later turn.

The same text now counts as more tokens

Haiku 5.5 uses the tokenizer introduced with Claude Opus 4.7. Anthropic: “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.” That is how 90% becomes 75%: the cut is measured against a token count that has grown. Simon Willison measured about 1.25× on his own prompts.

The window is nominally bigger, actually smaller: “1M tokens is roughly 555k words … on the current tokenizer”, while “models before it fit about 750k words in 1M tokens”. A max_tokens limit tuned for Haiku 4.5 “may cut off equivalent output”, because thinking tokens now count toward it. Images move the same way: a 2,000 × 1,500 pixel image costs about 2.5 times 4.5’s visual tokens.

Migration is not a model-ID swap

Anthropic’s release notes: “Code written for Claude Haiku 4.5 can break on Claude Haiku 5.5.” Manual extended thinking (budget_tokens), temperature, top_p, top_k and assistant prefill “can each return a 400 error”, and so does the old computer_20250124 tool (now computer_toolset_20260801 on the Claude API and Google Cloud, computer_20251124 on Bedrock). Replaying thinking blocks after a change to system, tools or earlier turns can also 400, and thinking blocks “work only in the account that produced them, or in an account linked to it” — which matters if you replay stored conversations through another key. On Bedrock, structured outputs are unavailable.

Haiku 4.5 has a date, and two status labels

The deprecations table lists claude-haiku-4-5-20251001 as Active, retirement “not sooner than October 15, 2026” — eight days after 5.5 launched. Haiku 5.5’s own commitment runs to 7 October 2027. The same page promises “at least 60 days’ notice before model retirement for publicly released models”, so a shutdown announced today lands in December.

Haiku 4.5’s overview page opens with Legacy and says migrate; the deprecations table calls the same model Active — two states under Anthropic’s own definitions.

The same comparison on OpenAI’s sheet

OpenAI’s pricing page puts gpt-6-luna at $0.10 / $0.50 short-context and $0.20 / $0.75 in its long-context column. Below the line they cost the same per token and Haiku 5.5 reports higher benchmarks (Terminal-Bench 4.0: 39.2% vs 16.4%; OSWorld 2.1: 72.4% vs 48.9%). Above it, Haiku 5.5 costs 2.5× Luna’s long-context input and 3.3× its output. Willison puts Luna’s step at 272,000 tokens.

Our arithmetic: a 120,000-token prompt with 2,000 output tokens costs $0.065 on Haiku 5.5, and the same request at 60,000 tokens costs $0.007 — doubling the prompt costs 9.3×. The same text on Haiku 4.5 (about 92,000 tokens on its tokenizer, flat $1 / $5) costs about $0.102. Haiku 5.5 is cheaper there by 36%, not 90%.

What we could not verify

• 90% and 75% are Anthropic’s. 90% is a traffic distribution for the previous Haiku; 75% is an average whose weighting is not published. We read both on the announcement page; the footnote behind 75% was truncated in our fetch.
• 1.25× is one person’s prompts. Anthropic publishes “approximately 30%” and says the increase depends on content. We ran no token counts.
• OpenAI’s pricing page returned HTTP 403 to us; its Luna rows reached us via a search crawl. The 272,000-token threshold is Willison’s, not a number we read from OpenAI.
• We ran no tests: no API calls, no cost measurements, no reproduction of the 400 errors.
• Prices are the Claude Platform’s. Bedrock and Google Cloud “set their own retirement schedules”.

If you are the one buying

1. Count your prompts on the tokenizer you will be billed on: text that fits 4.5’s counter is roughly 30% larger on 5.5’s, and the line is drawn on the new count.
2. Assume agent sessions cross the line, then pick the fix: compaction, a smaller stable block, or a model with flat 1M pricing — not more caching.
3. Recompute cost estimates rather than reuse them. Anthropic’s checklist says it outright: revisit max_tokens limits and cost estimates.
4. Read the cliff as a discount that shrinks, not a price rise: at $0.50 / $2.50, Haiku 5.5 is still below 4.5’s $1 / $5. What is being tested is the 90%.
5. If you are on Haiku 4.5, treat 15 October as a planning date, not a shutdown date: the commitment is “not sooner than”, and the 60-day notice rule is what turns an announcement into a deadline.

The same shape appears on a marketplace: Kimi K3 on Amazon Bedrock — one model, four rates.