Claude Haiku 5.5 Is Cheaper Per Token, Not Per Workflow

Last Updated: October 13, 2026

What Anthropic actually shipped on 7 October

Haiku 5.5 is a small model aimed squarely at the work that dominates agent pipelines by volume: classification, routing, extraction and subagent calls. Anthropic's model overview lists a 1M-token context window, 128k max output tokens (300k on the Batch API behind a beta header), a June 2026 knowledge cutoff, release on 7 October 2026 and a retirement date no sooner than 7 October 2027.

The pricing is tiered by prompt length. Anthropic's announcement puts input at $0.10/MTok and output at $0.50/MTok for prompts up to 100,000 tokens, rising to $0.50 and $2.50 above it, against Haiku 4.5's flat $1/$5. Cache reads drop to $0.01/MTok under the threshold, from $0.10 on Haiku 4.5. Anthropic's own framing of the headline saving is careful: 90% lower under 100k, 50% lower over it, and around 75% less to run on average once you account for the fact that the newer tokenizer consumes more tokens per task.

Benchmarks moved a long way for a Haiku-class model. On Anthropic's published results, Haiku 5.5 scores 1620 on GDPval-AA v2.1 against 735 for Haiku 4.5 and 1437 for GPT-6 Luna, and 72.4% on the OSWorld 2.1 offline subset against 15.7%. It also scores 39.2% on Terminal-Bench 4.0 where Sonnet 5.5 scores 70.6% — which is the number that should govern your routing decisions more than the GDPval jump does.

The tokenizer change quietly invalidates your existing cost model

Anthropic flagged this explicitly rather than burying it, and almost every roundup this week skipped it anyway. Haiku 5.5 uses the same newer tokenizer as Claude 4.7 and later models, so the same input text produces approximately 30% more tokens than on Claude Haiku 4.5, with the exact increase depending on content. The migration guide is blunt about the consequence: recount your prompts with model set to claude-haiku-5-5 rather than reusing counts measured on Haiku 4.5.

If your finance model was built from a token-usage export taken last quarter, every per-item figure in it is now understated by roughly a third on the input side. That matters less when the unit price falls 90% — but it matters a great deal for anything you sized in tokens rather than dollars: max_tokens ceilings, chunk sizes, context budgets, the threshold at which your router escalates to a bigger model, and the 100k price boundary itself.

Two secondary effects compound it. Images cost more: Haiku 5.5 uses the high-resolution image tier, and an image of 2,000 by 1,500 pixels costs about 2.5 times as many visual tokens as on Haiku 4.5. If your extraction pipeline feeds scanned invoices or screenshots, that is a direct multiplier on your highest-volume step. And in conversations, Haiku 5.5 keeps thinking blocks from all earlier assistant turns in context, where Haiku 4.5 kept only the latest turn's — so multi-turn agent loops carry more input tokens than the tokenizer change alone explains, unless you enable thinking block clearing.

Then there is adaptive thinking, which is on by default at medium effort. Thinking tokens bill as output, at the output rate, and count toward max_tokens. A classification step that previously returned a 40-token JSON object on Haiku 4.5 with thinking off can now return 40 tokens of answer behind several hundred tokens of reasoning. Anthropic's prompting guide notes that in its testing, telling the model in the prompt to answer directly did not stop it from thinking — you lower effort, or set thinking: {"type": "disabled"}, which works at low, medium and high only.

The 100k cliff is a pipeline design constraint, not a pricing footnote

Every other current Claude model prices the full 1M window at a flat rate. Haiku 5.5 does not. The pricing documentation spells out the mechanics that bite in production: a request's prompt length counts all of its input tokens, including cache reads and cache writes, and 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.

Read that twice if you run a long-lived agent with a large cached system prompt. A 90k-token cached preamble plus a 15k-token document is a 105k-token prompt. The whole thing prices at the upper tier, and the cache read you were counting on at $0.01/MTok bills at $0.05/MTok.

Arithmetic from the published Haiku 5.5 rates, uncached and excluding batch discounts. A 2% increase in prompt size across the boundary produces a 5x increase in cost. That is not a gradient your autoscaler will smooth out; it is a step function sitting in the middle of your highest-volume path.

Cheap classifiers get expensive in the exception queue

Here is the number that decides most demotion questions, and it has nothing to do with token price. In a real finance or support chain, the model is rarely the expensive component. The human who handles an exception is.

Take an invoice extraction pipeline at 10,000 documents a month. Classify the document type at roughly 1,200 input tokens and a few hundred output tokens including thinking; extract line items at roughly 3,000 input and 1,300 output. At Haiku 5.5's sub-threshold rates that is about $2.50 and $9.50 respectively, so roughly $12 of model spend for the month. The same extraction step on Sonnet 5.5 at $2/$10 would run about $190. Demoting saves you around $180 a month.

Now suppose the smaller model's exception rate on ambiguous documents rises from 4% to 7%. That is 300 additional items in a human review queue. At even a conservative $1.50 of fully loaded handling time per exception, the demotion has added $450 of cost to save $180. These figures are illustrative — substitute your own volumes and handling cost — but the shape holds across every chain we have priced: model spend is usually one to two orders of magnitude below exception-handling spend, so a three-point accuracy regression wipes out a 90% price cut.

The corollary is that per-token comparisons between Haiku 5.5 and, say, Google's Gemini 3.7 Flash — listed on Google Cloud's Agent Platform pricing page at introductory rates of $0.75/$3.75 per 1M tokens input/output through 31 December 2026 — are close to meaningless on their own. A 7.5x input price difference is $70 a month on the pipeline above. Which model produces fewer exceptions on your document distribution is worth hundreds. Run both against a labelled set and compare exception rates, not price sheets. We keep current rate and capability comparisons on our model coverage page for exactly this reason: the price is the easy half.

It's accurate enough that we'd trust it there, and fast and cheap enough that we can run it a lot.

— Alex Wang, Applied AI, Rogo, Anthropic launch page

Note the ordering in that sentence. Accuracy qualifies the step first; price determines how often you run it second. That is the right sequence.

The demotion test: which steps can drop to Haiku 5.5

Do not migrate a pipeline. Migrate one step, measure, then move to the next. The test below is what we run before changing a model inside a live n8n or custom orchestration layer, and it is deliberately boring.

Steps that reliably survive demotion: fixed-schema extraction, deduplication, language detection, intent classification with fewer than about twenty classes, compaction of long tool output before it reaches a larger model, and read-only database queries. Steps that usually do not: multi-document reconciliation, anything requiring a policy judgement, long agentic coding loops, and any step whose output is written to a system of record without a second check. We walk through this sequencing in more depth on our AI implementation work.

The migration checklist nobody reads until something 400s

Haiku 5.5 is not a drop-in model ID swap from Haiku 4.5. The official migration guide lists several changes that return hard errors, and these are the ones that surface as 400s in a production queue rather than as degraded output:

Two more that catch orchestration layers specifically. First, changing system, tools or earlier messages mid-conversation while sending thinking blocks back can return a 400 — keep conversations append-only. Any workflow node that rewrites the system prompt between turns, which is a common pattern in visual builders, needs rework. Second, changing the top-level effort value between requests invalidates the prompt cache for that conversation's messages. A router that dials effort up for hard items and down for easy ones will silently destroy your cache hit rate and your cost model with it; there is a per-message effort change in beta that preserves the cache.

If you store and replay conversations — common in audit-heavy workflows — note that thinking blocks work only in the account that produced them, or in an account linked to it.

Where this does not pay off yet

The honest version: if your pipeline is already stable on Haiku 4.5 and your monthly model spend is under a few hundred dollars, the migration is engineering time you will not recover. Seven breaking changes, a token re-baseline, a confidence-gate recalibration and a fresh eval run is a week of careful work. The payback on $200 a month of model spend is not there.

It also does not pay off if the step you want to demote is the one making the decision. Terminal-Bench 4.0 at 39.2% versus Sonnet 5.5's 70.6% is the clearest signal Anthropic published, and their own framing is that Sonnet and Opus remain the better choices for complex agentic coding while Haiku 5.5 suits narrowly scoped work that was previously cost-prohibitive.

And the broader constraint is not model price at all. A Teradata-commissioned survey of 1,000 senior technology and data leaders found 69% of C-suite executives say their organisation is already operating with agentic AI against 57% of VPs, that 40% of respondents say more than 40% of AI pilots pause before production because infrastructure is not ready for autonomy, and that 78% report difficulty creating a connected data foundation. None of those failures are priced in tokens. A cheaper model does not fix a pipeline that cannot reliably reach the right record at the moment of the decision. If that is where you are, fix the data path first — the model bill will still be the smallest line item when you get there. Examples of how we sequence that work are on our case studies page, and if you want a second opinion on a specific chain before you swap anything, start here.