Claude Haiku 5.5: What Anthropic's Cheapest Model Actually Is

Claude Haiku 5.5 is the smallest current first-party model — the tier above it is Claude Sonnet 5.5, and the model it replaces is Claude Haiku 4.5 — and it is the first Haiku that behaves like a large-context model: 1,000,000 tokens of input, up to 128,000 tokens of output, text and images in, text out. The vendor released it on October 7, 2026, and prices it from $0.10 per million input tokens and $0.50 per million output tokens. Those five facts are what most people searching this model's name want, and everything below is those facts with the details that change what you actually pay and what your code has to do.

Every figure here comes from one of two places, and each is named where it appears: the vendor's own published documentation, or an independent measurement by Artificial Analysis. Nothing on this page is carried over from another article on this blog, and where a widely repeated claim could not be re-established from the vendor, the page says so instead of repeating it.

Where Claude Haiku 5.5 sits in the line

The current first-party line is four models, and Haiku is the smallest of them. Positioned by the vendor's own descriptions of what each is for, and by the numbers the vendor publishes for each:

• Claude Haiku 5.5 — for high-volume, latency-sensitive work: classification, extraction, routing, subagent tasks. Fastest of the four. From $0.10 / $0.50 per million tokens. Default effort medium.

• Claude Sonnet 5.5 — the middle tier, for general production work. Fast. $2 / $10 per million tokens. Default effort high.

• Claude Opus 5.5 — for long-running agentic coding and knowledge work. Moderate latency. $4 / $20 per million tokens. Default effort medium.

• Claude Fable 5.1 — the most capable widely released model, for demanding reasoning and long-horizon agents. Slowest. $10 / $50 per million tokens. Default effort high.

Against the generation it replaces, the move is larger than a version bump. Claude Haiku 4.5 is a 200,000-token context model with a 64,000-token output limit and a flat $1 / $5 per million tokens; Claude Haiku 5.5 is a 1M-context, 128K-output model that starts at a tenth of that rate. The vendor commits not to retire Haiku 5.5 before October 7, 2027 — the same two-year floor it gives the rest of the current line — which matters more than usual for a small model, because small models are the ones people wire into long-lived pipelines and forget about.

The pricing, and the step at 100,000 tokens

The vendor's published price for Claude Haiku 5.5 is not one price. It is two, split at a prompt length, and the split is the single most consequential thing on this page for anyone estimating cost.

• Prompt up to 100,000 tokens — input $0.10 per million, output $0.50 per million, five-minute cache write $0.125, one-hour cache write $0.20, cache read $0.01.

• Prompt over 100,000 tokens — input $0.50 per million, output $2.50 per million, five-minute cache write $0.625, one-hour cache write $1.00, cache read $0.05.

• Batch API — 50% off input and output at either tier, and up to 300,000 output tokens with the output-300k-2026-03-24 beta header, against 128,000 on the synchronous Messages API.

The vendor describes the second tier plainly in its pricing documentation: a request whose prompt exceeds 100,000 tokens pays the higher prices, and the prompt length counted is all of its input tokens including cache reads and cache writes. Each request is priced on its own — a long request with a partial cache hit still pays the higher tier — and requests that already ran keep the prices they were billed at.

The vendor also publishes the arithmetic behind its headline savings claim, and the arithmetic is more interesting than the headline. Haiku 5.5's rate card is 90% below Haiku 4.5's for prompts up to 100,000 tokens and 50% below it above that; the vendor says roughly 90% of requests to the previous Haiku model fell under the threshold. But it also says the new model uses the same tokenizer as Claude 4.7 and later, so the same text counts as approximately 30% more tokens than on Haiku 4.5 — which is why the vendor's own summary is "around 75% less to run", not 90%. That 75% is a vendor-reported average, not a re-measured benchmark, and it depends on a workload mix you may not have.

The practical version: if your prompts are short and your cache hit rate is good, the cut is real and large. If your prompts run long, the ceiling is what changed most — you can now send a 1M-token prompt at all, and you pay five times the short-prompt rate to do it.

Two things that circulate with this model are worth separating from the vendor's own material. The first is the "75% cheaper" figure, which is Anthropic's and carries the tokenizer caveat above; treat it as the vendor's estimate rather than a measured saving. The second is the claim that Claude Haiku 5.5 was aimed at undercutting Chinese flash-tier models. That is not something Anthropic's material says. Every comparison Anthropic publishes for this model is against Anthropic's own previous Haiku — Haiku 4.5 — and the positioning line is about high-volume, latency-sensitive workloads, not about a competitor's price. Treat the undercut framing as unestablished unless someone shows you a primary source for it.

Context window, output limit, modalities

Anthropic's specification page for Claude Haiku 5.5 gives the following, and each of these is the vendor's own published figure rather than an inference from the model's name:

• Context window — 1,000,000 tokens.

• Maximum output — 128,000 tokens on the synchronous Messages API; 300,000 on the Batch API with the extended-output beta header.

• Modalities — text and images in, text out. There is no audio or video input path, and the output is text; Anthropic's capability table states "text and images → text". Artificial Analysis independently confirms the same input/output pair on its own page for the model.

• Thinking — adaptive, on by default, steered by the effort parameter, whose default on this model is medium. This is the first Haiku with an effort dial; Haiku 4.5 used a manual thinking budget instead.

• Knowledge cutoff — June 2026, both the reliable cutoff and the broader training-data cutoff.

Two consequences of adaptive thinking are easy to miss and both are documented by the vendor as changes from Haiku 4.5. A response can now begin with thinking blocks, so code that reads the first content block as the answer will break — select blocks by type. And thinking blocks from earlier turns stay in context and are billed as input tokens, where Haiku 4.5 kept only the newest turn's, so multi-turn conversations carry more input than the tokenizer change alone would explain.

What Anthropic claims — and what it actually measured

Everything in this section is vendor-reported. Anthropic published the following figures for Claude Haiku 5.5 in its release material, each against Claude Haiku 4.5 as the comparison point, and none of them has been reproduced here:

• OSWorld 2.1, computer use, offline subset — 72.4% for Claude Haiku 5.5 against 15.7% for Claude Haiku 4.5.

• Terminal-Bench 4.0, agentic coding — 39.2% against 0.0%.

• Humanity's Last Exam, multidisciplinary reasoning — 45.9% without tools and 57.4% with tools, against 10.2% and 18.7%.

• GDPval-AA v2.1, knowledge work — 1620 against 735.

• AA-Briefcase v1.1 — 1578 against 614.

• FrontierCode 1.1, main — 46.4%.

• Chartography, visual reasoning — 46.4% without tools against 6.4%.

The vendor's own framing of these numbers is worth quoting in substance: Anthropic says Claude Sonnet 5.5 and Claude Opus 5.5 "remain better choices for complex agentic coding tasks like those measured by Terminal-Bench 4.0", and positions Haiku 5.5 for narrowly scoped tasks that would otherwise be too expensive to run at all. Anthropic has not published a head-to-head run of Haiku 5.5 against Sonnet 5.5 at matched effort on a matched harness, so nothing here should be read as a ranking between them; the vendor's benchmark set is its own and is labelled as such.

The independent picture comes from Artificial Analysis, which ran the model at its Max effort configuration and reports the following on its own page for Claude Haiku 5.5: an Intelligence Index score of 43, described there as well above the median of 13 for comparable models; a measured 236 output tokens per second against a 106-token median for similar models; a cost of $0.21 per completed Intelligence Index task; and a notable verbosity flag — 440M tokens generated during the Index run against a median of 95M. The board also records the model's pricing as $0.10 per million input tokens against a $0.25 median, its text-and-image input, and its 1M-token context window, which is the independent confirmation of the modality and window claims above. Those are Artificial Analysis's measurements, not Anthropic's, and they were taken on the Max-effort configuration of the model — the board notes a non-reasoning configuration may exist, and that figures for it may differ. The board does not publish a measurement date on the page.

How you get at Claude Haiku 5.5

The vendor shipped Claude Haiku 5.5 on the widest first-day surface it has given a Haiku model. The model IDs, as published:

• Claude API — claude-haiku-5-5.

• Bedrock — anthropic.claude-haiku-5-5.

• Vertex AI, Microsoft Foundry and Claude Platform on AWS — claude-haiku-5-5 on each.

The wide surface is the point: many teams will meet this model through a cloud contract they already have rather than a new vendor relationship, and the model ID is the same string everywhere except Bedrock's prefixed form. Anthropic has not published a free tier, free credits or a trial for Claude Haiku 5.5, and nothing on this page should be read as one.

On OrcaRouter's side, one thing has to be said plainly rather than left to implication: our live catalogue does not serve Claude Haiku 5.5 today. Checked against the live model list, which carries 207 models, the Anthropic entries run through Claude Opus 5.5 and Claude Sonnet 5.5 but stop at Claude Haiku 4.5 for the small tier — so if you are routing Haiku-class traffic through us now, you are on the previous generation. If Anthropic's price for the new model changes, that is the one case where a passed-through price moves on our side the same day, because we pass the provider list price through and bill with no markup; but that is a statement about how our pricing works, not a claim that this particular model is available. When it appears, it will appear in the same catalogue.

What this model is, in one paragraph

Claude Haiku 5.5 is a 1M-context, 128K-output, text-and-image small model that Anthropic sells from $0.10 and $0.50 per million tokens, steps to $0.50 and $2.50 above 100,000 input tokens, and steers with a single effort dial rather than the sampling knobs earlier Haiku models accepted. Its measured speed and its measured Intelligence Index score both sit well above the median for models in its price band, its verbosity sits the other way, and the vendor's benchmark claims are the vendor's. If you are choosing between it and something else in the tier, the two questions that decide it are whether your prompts fit under the 100,000-token step and whether you can live with a 50% batch discount instead of a synchronous call.