Claude Haiku 5.5: 90% Cheaper Than Haiku 4.5 — With Two Catches

Anthropic released Claude Haiku 5.5 on October 7, and the headline number is hard to ignore: $0.10 per million input tokens, a 90% cut from Haiku 4.5. Small models rarely make news. This one should, because small models are where the volume is — every summarizer, classifier and subagent in your pipeline bills by the token, thousands of times a day. But the sticker price is not the invoice. There are two catches in the fine print, and Anthropic, to its credit, printed both of them.

The quick answer

• Price, prompts up to 100K tokens: $0.10 input / $0.50 output per million tokens
• Price, prompts over 100K tokens: $0.50 / $2.50 — still 50% below Haiku 4.5
• Batch API: half again — $0.05 / $0.25
• Anthropic's own blended estimate: around 75% cheaper to run than Haiku 4.5
• Available now on: Claude Platform, AWS, Google Cloud and Microsoft Azure

The numbers

Two smaller changes ride along: the shortest cacheable prompt drops from 4,096 tokens to 512, so even short system prompts can now be cached, and the same launch halved Sonnet 5.5's cache-read price to $0.10 per million tokens. The pricing structure itself, though, is where the story is.

Catch #1: the tokenizer quietly eats part of the discount

Haiku 5.5 uses the newer tokenizer already running in Sonnet 5.5 and Opus 5.5. The same text counts as roughly 30% more tokens than it did on Haiku 4.5. Do the arithmetic and the 90% headline cut becomes roughly an 87% effective cut on short-prompt input — still enormous, but not the number on the poster. On long prompts the nominal 50% cut shrinks to around 35%.

This is exactly why Anthropic's own summary says "around 75% less to run" instead of 90%: the blended figure folds in the tokenizer change and the real request mix, and it is the number to budget against. Practical consequence: before migrating, recount your prompts. Every token count, limit and cost estimate you measured on Haiku 4.5 is now off by about a third.

Catch #2: the 100K cliff

The cheap tier is set per request by prompt length, and it is a cliff, not a slope. A 400K-token prompt does not pay $0.10 on its first 100K tokens and $0.50 on the remainder — the entire request bills at the long-prompt tier. Haiku 5.5 advertises a 1M-token context window, but the budget price only patrols the first tenth of it.

Anthropic's defence is reasonable: around 90% of requests to Haiku 4.5 were under 100K tokens, so the cheap tier covers how the model is actually used. If your workload lives in the other 10% — long documents, big codebases, fat context — your discount is 50%, not 90%. Plan on that basis and the pricing is still attractive; plan on the headline and your first invoice will educate you.

It is not just cheaper — it is much better

A price cut on a weaker model would be a footnote. In Anthropic's published tests, Haiku 5.5 scored 1,620 Elo on GDPval-AA v2.1, a benchmark of professional knowledge work across 44 occupations — against 735 for Haiku 4.5 and 1,437 for OpenAI's GPT-6 Luna. On OSWorld 2.1, which measures how well an agent operates a computer, it reached 72.4% on the offline subset, up from 15.7% for its predecessor. On Terminal-Bench 4.0, an agentic coding test, it scored 39.2% where Haiku 4.5 scored zero.

Independent testing is more sober and still positive: Artificial Analysis scores the model between 34 and 43 on its Intelligence Index depending on effort — and Haiku has an adjustable effort setting for the first time, so you choose where on that range each request sits, and what it costs (roughly $0.05 to $0.21 per benchmark task in their measurements). Two grains of salt, both worth taking: most headline benchmarks are the vendor's own, and Anthropic itself steers complex agentic coding to Sonnet 5.5 and Opus 5.5. Haiku's job is the high-volume work underneath the flagship's job.

Who should switch, and who should not

If your bill comes from classification, summarization, extraction, routing or subagent grunt work, this is close to a free upgrade: the same API shape, a dramatically lower bill, and better scores. Migrate — but recount tokens first, and judge the move on cost per completed task, not cost per token, because accuracy and retries move the real number. If your workload is a handful of hard, long-context reasoning calls, the math is thinner and the flagships remain the right tool.

Either way, the $1 small model at Anthropic is finished, and every competing lab now has to answer a $0.10 price point. If you are rebuilding a stack around cheaper inference, our running guide to the best new AI tools worth trying in 2026 is the place to start.

Frequently asked questions

Is Claude Haiku 5.5 really 90% cheaper?

Per token, on prompts up to 100K tokens, yes: $0.10/$0.50 against Haiku 4.5's $1/$5. In practice the new tokenizer counts about 30% more tokens for the same text, so Anthropic's own blended estimate is around 75% lower running costs.

What is the 100K token cliff?

Pricing is set per request by prompt length. Cross 100K tokens and the whole request bills at the higher tier ($0.50 input / $2.50 output per million) — there is no blended rate for the portion under the line.

Claude Haiku 5.5 or GPT-6 Luna?

On Anthropic's published tests, Haiku 5.5 edges Luna on knowledge work (1,620 vs 1,437 Elo on GDPval-AA v2.1) and computer use. Treat vendor benchmarks as a starting point, not a verdict: measure cost per completed task on your own workload.

Where can I use Claude Haiku 5.5?

On the Claude Platform (API model name claude-haiku-5-5), Amazon Web Services, Google Cloud and Microsoft Azure.

Key takeaways

• Haiku 5.5 launched October 7 at $0.10/$0.50 per million tokens for prompts up to 100K — 90% below Haiku 4.5
• Catch one: the new tokenizer counts ~30% more tokens, so budget on Anthropic's ~75% blended saving, not 90%
• Catch two: past 100K tokens the entire request bills at the higher tier — there is no blended rate
• The benchmarks jumped too: GDPval-AA 1,620 vs 735, OSWorld 72.4% vs 15.7% (Anthropic's published tests)
• Batch API halves prices again to $0.05/$0.25 — currently the cheapest mainstream route for bulk work

Prices verified against Anthropic's official pricing page, October 2026. Benchmarks are Anthropic's published figures unless marked as independent.