Claude Haiku 5.5: Adaptive Thinking by Default, a Stricter API, and a Price That Steps at 100,000 Tokens

Claude Haiku 5.5 went live on October 7, 2026 — two days before this writing — and most of the launch coverage has been about the score. The score is real: on the independent Artificial Analysis Intelligence Index the model lands at 43, second of 182 entries on that board, which puts A​nthropic's cheapest current tier above most vendors' flagships. The part that will change your afternoon is quieter. The API got stricter, the thinking became adaptive by default, and both changes are the kind that break existing code rather than fill a launch post.

What follows is the working summary: what actually shipped, what the API no longer accepts, what the pricing does at volume, and which of the vendor's numbers deserve an arm's length.

What shipped on October 7

Claude Haiku 5.5 is the model ID claude-haiku-5-5, the direct replacement for Claude Haiku 4.5, and it is a text-and-image model: image in, text out, no audio or video path. It ships with a 1,000,000-token context window and a 128,000-token maximum response, extendable to 300,000 tokens on the Batch API through the output-300k-2026-03-24 beta. The knowledge cutoff is June 2026, and A​nthropic commits to not retiring the model before October 7, 2027 — a two-year floor, which matters more than usual for a small tier that people wire into long-lived pipelines.

Availability is broad from day one: the Claude API itself, Amazon Bedrock, G​oogle Cloud, Microsoft Foundry, and Claude Platform on AWS. That is the widest first-day surface A​nthropic has given a Haiku model, and it is the reason a lot of teams will meet this model through an existing cloud contract rather than a new one.

The API is stricter than the model it replaces

This is the section to read before you point an existing Haiku 4.5 client at the new ID. Claude Haiku 5.5 rejects several request shapes that older Claude models accepted, and it rejects them with a 400 rather than a warning:

• Sampling parameters — setting temperature, top_p or top_k to a non-default value returns an error. The knobs are gone, not ignored.

• Assistant prefill — sending a partial assistant turn for the model to continue is no longer supported on this model.

• Manual extended thinking — passing budget_tokens to size a thinking budget returns an error, because thinking is now governed by the effort dial described below.

• Computer use — the tool requires the computer_toolset_20260801 toolset version; the older toolset identifier is not accepted.

Any one of those is a one-line fix, but they are the kind of one-line fix you discover in production at 2am unless you read the migration note first. The pattern is consistent: A​nthropic has removed the per-request improvisation and moved those decisions to model-level configuration. For a fleet calling one model ID across many services, that is a real simplification. For a service that was relying on a temperature sweep or a hand-sized thinking budget, it is a rewrite of the calling convention.

Adaptive thinking, and a dial from Low to Max

The replacement for all of that is adaptive thinking, which is on by default at a default effort of medium. The model decides, per request, how much internal reasoning the task needs, and you steer it with a single effort setting that runs from Low through Medium, High and Max.

The practical consequence is the one people underrate: you can no longer make the model cheap by turning thinking off. Thinking is the product. What you control is how hard it tries, and the effort setting is a cost dial as much as a quality dial — Low on a classification job and Max on a debugging job is the intended shape of a production deployment, not two different models.

It also reframes what a "small model" means. Historically you chose a Haiku-class model by accepting that it would fail the hard cases and routing those elsewhere. Claude Haiku 5.5 at Max effort is meant to close some of that gap itself, which changes the routing decision from "can this model do the task" to "is this task worth Max effort".

Pricing: flat until 100,000 tokens, then five times

Per million tokens in US dollars, A​nthropic's published rates are $0.10 in and $0.50 out up to 100,000 tokens of input, stepping to $0.50 and $2.50 above that line. Caching is priced separately: a five-minute cache write is $0.125 and $0.625, a one-hour write is $0.20 and $1.00, and a cache read is $0.01 and $0.05. Batch inference is half price.

The step at 100,000 tokens is the number to build into a spreadsheet rather than a footnote. A pipeline that runs comfortably at 90,000-token prompts pays the first tier; the same pipeline with a retrieval setting nudged one notch higher pays five times the input rate on every token in the request, not just the tokens past the line. Long-context work that used to be a rounding error becomes the dominant cost of the call.

Read as a comparison to the model it replaces, the headline is real: A​nthropic states Claude Haiku 5.5 runs about 75% cheaper on average than Claude Haiku 4.5, and about 90% cheaper specifically for prompts under 100,000 tokens. That is a vendor figure and it is stated as one. The mechanism behind it is not only the rate card — A​nthropic's own documentation notes that Claude Haiku 5.5 uses the newer tokenizer shared with Claude 4.7 and later, so the same text tokenizes to roughly 30% more tokens than it did on the old model. A per-token cut and a per-token inflation pull in opposite directions, and the 90% figure is the net after both.

The benchmarks, as A​nthropic reports them

Every number in this section is vendor-reported and unreproduced by a third party. A​nthropic's comparison set runs Claude Haiku 5.5 against Claude Haiku 4.5, GPT-6 Luna, and its own Claude Sonnet 5.5:

• GDPval-AA v2.1 — 1620 for Claude Haiku 5.5, against 735 for Claude Haiku 4.5, 1437 for GPT-6 Luna and 1840 for Claude Sonnet 5.5.

• AA-Briefcase v1.1 — 1578, against 614 for Claude Haiku 4.5 and 1824 for Claude Sonnet 5.5.

• OSWorld 2.1, offline — 72.4%, against 15.7% for Claude Haiku 4.5, 48.9% for GPT-6 Luna and 83.9% for Claude Sonnet 5.5.

• Humanity's Last Exam — 45.9% without tools and 57.4% with them, against 10.2% and 18.7% for Claude Haiku 4.5, and 56.9% and 64.5% for Claude Sonnet 5.5.

• Terminal-Bench 4.0 — 39.2%, against 0.0% for Claude Haiku 4.5, 16.4% for GPT-6 Luna and 70.6% for Claude Sonnet 5.5.

• FrontierCode 1.1, Main — 46.4, against 42.4 for the same model at low effort and 52.1 at Xhigh.

Two things are worth noticing about that table even before anyone reproduces it. First, the jump over Claude Haiku 4.5 is not incremental on agentic and computer-use tasks — 15.7% to 72.4% on OSWorld is a different class of result, and the Terminal-Bench row moves from a literal zero. Second, the vendor's own effort sweep shows the dial doing real work: the same model scores 42.4 at low effort and 52.1 at Xhigh on FrontierCode. That is the adaptive-thinking claim restated as a number, and it is the reason the effort setting belongs in your cost model rather than your defaults.

What the independent board adds

The neutral check available today is Artificial Analysis, which measured Claude Haiku 5.5 at Max effort on its Intelligence Index and returned 43 — second of 182 models, with a blended cost of $0.08 per million tokens and a measured $0.21 per finished Intelligence Index task. Context is confirmed at 1.0M tokens, and throughput came in at 240.4 output tokens per second, against a 110 tokens-per-second median across the board.

The same run flags one thing the vendor's material does not: token appetite. Artificial Analysis recorded 440 million output tokens while running the Index, against a 100-million median, and calls the model verbose. On an input-and-output-priced model, a model that writes four times the median number of tokens per answer is effectively more expensive than its rate card suggests, and the $0.21-per-task figure already accounts for that. It is a fair number to hold alongside the $0.10 sticker rather than instead of it.

A​nthropic also shipped a companion change on the same day: cache reads on Claude Sonnet 5.5 were halved, from $0.20 to $0.10 per million tokens, and Max and Team plans gained monthly API credits ($100 on Max 5x, $200 on Max 20x, up to $500 pooled on Team). Neither changes Claude Haiku 5.5's price, but both shift where the cheap tier sits relative to the mid tier — a Halved Sonnet cache read narrows the gap a long-context pipeline sees between the two.

Routing a small tier that is this good

Claude Haiku 5.5 itself is not on the OrcaRouter catalogue yet, and it is better to say that plainly than to leave the availability line vague. It is reachable today through A​nthropic's own API and the cloud platforms above.

What the model does change is the shape of a routing decision, and that is where a router earns its place. If a small tier at Index 43 is now credible on work you used to send to a mid tier, the useful configuration is a three-leg route — a cheap leg for the easy majority, this tier for the hard middle, and a top-of-range model for the tail — and on OrcaRouter that is a routing DSL expression rather than a project. Claude Haiku 4.5, Claude Sonnet 5.5 and Claude Opus 5.5 are all on the catalogue today, so the A​nthropic legs of such a route can be built and tested now; the small-tier leg opens the day Claude Haiku 5.5 is added. One key, one SDK, automatic failover underneath, and provider list prices passed through at 0% markup, so a vendor cut on any leg is live on ours the same day.

What to watch

The open questions are the two the launch material cannot answer. The first is whether the 100,000-token cliff survives contact with real retrieval pipelines, or whether it pushes teams back under the line the way it is designed to. The second is whether adaptive thinking at Max effort holds up on long agentic runs, where the vendor's own effort sweep — 42.4 at low, 52.1 at Xhigh on FrontierCode — suggests the quality is there but says nothing yet about the token cost of getting it. Both will be answered by the first wave of independent agentic evaluations, not by the launch page.