How much does Claude Haiku 5.5 cost per million tokens?

For prompts up to 100,000 tokens it is $0.10 per million input tokens and $0.50 per million output. Above that threshold the same request costs $0.50 input and $2.50 output. Cache hits are $0.01 per million on the lower tier and $0.05 on the higher one.

Why is Claude Haiku 5.5 pricing listed as 'from $0.10'?

Because it is the only current Claude model priced by prompt length. Every other model from the 4.6 generation on bills the full one-million-token window at one flat rate. Haiku 5.5 has two rate tiers, split at a 100,000-token prompt, so the published table shows the lower one as a starting price.

Do cached tokens count toward the 100,000-token threshold?

Yes. Anthropic counts all input tokens in a request toward prompt length, including cache reads and cache writes. A request can therefore cross into the higher tier on cached context alone, and it pays the higher prices on every token even though most of the prompt was a cache hit.

Is Claude Haiku 5.5 really 75% cheaper than Haiku 4.5?

That is an average, not a rate cut. On sticker rates it is 90% cheaper for prompts up to 100,000 tokens and 50% cheaper above. Anthropic says about 90% of Haiku 4.5 requests fell into the lower band, and the blended figure also accounts for a new tokenizer that uses slightly more tokens per task.

Should I use Claude Haiku 5.5 or Sonnet 5.5 for coding?

Sonnet 5.5, for anything agentic. Haiku 5.5 scores 39.2% on Terminal-Bench 4.0 against Sonnet 5.5's 70.6%. Anthropic positions Haiku 5.5 for narrowly scoped, high-volume work such as classification, extraction and routing, or as a subagent under a larger model.

Does a cheaper small model change how technical interviews work?

It lowers the cost of the work candidates used to be tested on. Classification, extraction and boilerplate now run for cents, so interviews shift toward judgement: reading a diff, choosing which model a task deserves, and explaining why a cost or latency tradeoff is the right one.

Claude Haiku 5.5 Pricing: The 100K Token Cliff

Claude Haiku 5.5 pricing starts at $0.10 per million input tokens and $0.50 per million output, which makes it the cheapest model Anthropic has ever shipped. It is also the only model in the current lineup whose rate card changes depending on how long your prompt is. Cross 100,000 tokens and every number on that row multiplies by five.

That detail is missing from most of the pages currently ranking for this model, and it is the one that shows up on the invoice. Here is the full rate card, where the cliff bites, and what the "75% cheaper" headline actually measures.

What Anthropic shipped on 7 October

Anthropic released Claude Haiku 5.5 on 7 October 2026, calling it the cheapest, fastest, and most capable small model it has ever released. The API ID is claude-haiku-5-5, and it is available on the Claude API, Amazon Web Services, Google Cloud and Microsoft Azure.

The spec sheet is not a cut-down one. Per the published model overview, Haiku 5.5 carries the same 1M-token context window and 128K max output as Opus 5.5 and Sonnet 5.5, the same June 2026 knowledge cutoff, and adaptive thinking with a default effort of medium. It is the first Haiku-class model with an adjustable effort setting, running from low through to max. Anthropic commits to keeping it available until at least 7 October 2027.

It is also the fastest model in the lineup at standard speed — though Anthropic notes the Opus models in Fast Mode still beat it.

Claude Haiku 5.5 pricing, rate by rate

Two tiers, split at a 100,000-token prompt. All figures per million tokens, from the Claude API pricing page.

Prompts up to 100,000 tokens

Input: $0.10
Output: $0.50
5-minute cache writes: $0.125
1-hour cache writes: $0.20
Cache hits and refreshes: $0.01

Prompts over 100,000 tokens

Input: $0.50
Output: $2.50
5-minute cache writes: $0.625
1-hour cache writes: $1.00
Cache hits and refreshes: $0.05

Every line is exactly five times the lower tier. There is no taper and no blend — one token over the line reprices the whole request.

Batch processing halves both tiers

The Message Batches API is 50% off, as it is across the lineup. That puts Haiku 5.5 at $0.05 input and $0.25 output under 100K, and $0.25 input and $1.25 output above it. For anything you can run asynchronously — nightly scoring, bulk tagging, re-indexing — batch is close to free at this scale.

The 100K cliff is unique to Haiku 5.5

This is the part worth internalising. Anthropic's pricing documentation is explicit that Claude 4.6 and later models include the full 1M-token context window at standard pricing — except Claude Haiku 5.5. On Opus 5.5 or Sonnet 5.5, a 900,000-token request is billed at the same per-token rate as a 9,000-token one. On Haiku 5.5 it is not.

The rule that costs people money: prompt length counts all input tokens in a request, including cache reads and cache writes. A request can cross into the five-times tier on cached context alone — and once it crosses, it pays the higher rate on every token, including the cached ones. Keep a Haiku 5.5 prompt under 100,000 tokens and you are paying $0.10. Let it drift to 101,000 and you are paying $0.50 on all of it.

Cache reads count toward the threshold

This inverts the usual intuition about caching. Everywhere else, a cache hit is the cheap path. Here a large cached prefix is also what pushes you over the line, and the discount it earns is applied at the higher tier. Caching still pays — $0.05 against $0.50 is a 10x saving — but it no longer protects you from the cliff.

Every request is priced on its own

There is no averaging across a session. A conversation that sits at 95K for twenty turns and spikes to 105K on the twenty-first pays the lower rate twenty times and the higher rate once. Earlier requests keep the prices they were billed at, so a single long turn does not retroactively reprice the thread.

US-only inference multiplies the higher tier too

If you set inference_geo to US-only, Claude 4.6 and later models carry a 1.1x multiplier on every pricing category. On Haiku 5.5 that multiplier applies to the over-100K prices as well, so the two surcharges compound. Global routing is the default and uses standard pricing.

Practically, the fix is architectural rather than clever: if you are feeding a big document to Haiku 5.5, chunk it. Two 60K requests at the lower tier cost less than one 120K request at the higher one, for the same tokens.

What "75% cheaper" actually measures

Anthropic says the average cost to run Haiku 5.5 is about 75% lower than Haiku 4.5. That is a blended figure, not a rate cut, and the footnote on the announcement page explains the blend: it is 90% cheaper for requests up to 100,000 tokens and 50% cheaper above, and about 90% of Haiku 4.5 requests fell into the first group.

The same footnote adds a detail most summaries drop — the calculation already accounts for Haiku 5.5's updated tokenizer, which uses slightly more tokens per task. So the 75% is an honest number, but it describes Anthropic's traffic mix rather than yours. If your workload is long-document summarisation, your realised saving is closer to 50%. If it is short classification calls, it is closer to 90%.

Haiku 4.5, for reference, was a flat $1 input and $5 output with no tiers at all.

Where Haiku 5.5 lands on capability

Cheap is only interesting if the model is good enough for the job. Anthropic published a comparison against Haiku 4.5, GPT-6 Luna and Sonnet 5.5:

OSWorld 2.1 (offline subset): 72.4% against Haiku 4.5's 15.7% and GPT-6 Luna's 48.9%. Sonnet 5.5 leads at 83.9%.
GDPval-AA v2.1 (Elo): 1620 against 735 for Haiku 4.5 and 1437 for GPT-6 Luna. Sonnet 5.5 sits at 1840.
Humanity's Last Exam (with tools): 57.4% against Haiku 4.5's 18.7%, closing on Sonnet 5.5's 64.5%.
Terminal-Bench 4.0: 39.2%, against 0.0% for Haiku 4.5 and 70.6% for Sonnet 5.5.

The generational jump over Haiku 4.5 is enormous — a model that scored zero on Terminal-Bench now clears 39%. But the Terminal-Bench gap to Sonnet 5.5 is the honest signal, and Anthropic says so directly: Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding. Haiku 5.5 is positioned for narrowly scoped work — summaries, compaction, database queries, classification — and as a subagent underneath a larger model.

Anthropic also cut Sonnet 5.5 cache reads by 50% in the same announcement, from $0.20 to $0.10 per million, which it says makes Sonnet 5.5 roughly 20% cheaper on most agentic tasks. If you read our comparison of Sonnet 5.5 and Opus 5.5 for coding, that changes the arithmetic in Sonnet's favour.

How we think about a cheap fast model at Taqari

We run a lot of small, repetitive language work that is nothing like agentic coding: classifying what a candidate's answer was actually about, extracting the structured fields out of a transcript, routing a session to the right rubric, compacting a long interview into something the scorer can read. None of that needs a frontier model, and all of it is latency-sensitive because a candidate is sitting there waiting.

Those calls are also naturally short. A single answer plus a rubric fragment is nowhere near 100,000 tokens, which is exactly the shape the lower tier rewards. The place we would have to be careful is compaction over a full session transcript — that is the one path where a prompt grows without anyone deciding it should, and it is the path worth putting a token budget on.

What we would not do is move scoring itself onto a small model. We wrote about building an LLM-as-judge rubric that scores reproducibly, and the thing that makes those scores stable is a model that reasons carefully about an ambiguous answer. That is Sonnet or Opus work. The cost case for Haiku is real, but it belongs on the plumbing, not on the judgement.

What this changes if you are preparing for interviews

Every release like this moves the line between what a person is paid to do and what a model does for a fraction of a cent. Classification, extraction, summarisation and boilerplate are now genuinely cheap — a million tokens of output for fifty cents. Writing that code from memory is not the skill being tested any more.

What has become more valuable is the judgement around it: knowing that a 120K prompt should be two 60K prompts, that cache reads count toward a threshold, that a benchmark average describes someone else's traffic. Those are the decisions that separate an engineer who ships a working system from one who ships an expensive one, and they are increasingly what a good interviewer probes for.

If you want to practise explaining tradeoffs like these out loud rather than just reading about them, that is what Taqari's free AI mock interviews are for. And if you are pricing a real workload, our breakdown of what Claude Opus 5.5 actually costs to run covers the other end of the same lineup.