claude haiku 5.5 vs sonnet 5.5: which jobs should a small business give the cheap model?

elisabeth hitz · published october 13, 2026 · updated october 13, 2026 · 8 min read

anthropic released claude haiku 5.5 on october 7, 2026, nine days after sonnet 5.5. the headline is the price. the part that decides your bill is one line in the pricing table that most people will not read.

give claude haiku 5.5 the short, repetitive, high-volume jobs: summaries, classification, lookups, first drafts of routine replies. keep sonnet 5.5 for anything long, multi-step, or expensive to get wrong. haiku 5.5 costs a twentieth of sonnet 5.5 per token on prompts under 100,000 tokens, and five times its own base price above that line.

that last sentence is the whole post. a small business does not need the cheapest model. it needs to know which of its jobs are haiku-shaped, and which of those will quietly cross the 100,000-token line and get billed at the other price. so here is what anthropic released, what the three "cheaper" numbers actually compare, where the line sits, and a four-step test to run on your own work.

what anthropic actually released

haiku 5.5 is the small, fast tier of the claude 5.5 family, released on october 7, 2026 for high-volume, cost-sensitive work. anthropic describes it as its cheapest, fastest, and most capable small model so far, built for quick and repetitive workloads such as summaries, compactions, database queries, and classification, as a subagent next to opus 5.5 and sonnet 5.5, and for speed-sensitive jobs like live customer support and browser use (anthropic, oct 7 2026).

two details matter more for a business than the benchmarks. first, it is in the consumer app: free, pro, max, team, and enterprise users can select haiku 5.5 on claude.ai, on web, ios, and android (anthropic claude haiku page, accessed oct 13 2026). on the api it is claude-haiku-5-5, and it is also on amazon web services, google cloud, and microsoft azure (anthropic, oct 7 2026). second, it is the first haiku-class model with an adjustable effort setting (anthropic, oct 7 2026), so the dial i covered in which claude effort level should you use now applies to the cheap model too.

anthropic also draws the line itself. sonnet 5.5 and opus 5.5 remain the better choice for complex agentic coding, and haiku 5.5 is best suited to narrowly scoped tasks that were too expensive to automate before (anthropic, oct 7 2026). that sentence is the routing rule. everything below is how to apply it.

the numbers, side by side

on list price, haiku 5.5 is a twentieth of sonnet 5.5 under 100,000 prompt tokens, and on complex multi-step work it scores a little over half of sonnet. every figure below is from anthropic's own announcement table, which compares haiku 5.5 with haiku 4.5 and sonnet 5.5.

opus 5.5 is not in anthropic's haiku table, so it is not in mine. for the record, anthropic's pricing page lists it at $4 input and $20 output per million tokens (anthropic pricing page, accessed oct 13 2026), and i compared it with sonnet in claude sonnet 5.5 vs opus 5.5. the gdpval-aa numbers there are from an earlier version of the benchmark, which is why they do not match the v2.1 scores above.

three "cheaper" numbers that mean three different things

there are three savings figures attached to this launch, and all three compare haiku 5.5 with haiku 4.5, not with sonnet 5.5. keep them apart or you will budget off the wrong one.

1. 90% is the list-price cut for prompts up to 100,000 tokens. $0.10 and $0.50 per million tokens against haiku 4.5's $1 and $5 (anthropic, oct 7 2026).
2. 50% is the cut above 100,000 tokens. $0.50 and $2.50 against the same $1 and $5 (anthropic, oct 7 2026).
3. around 75% less is anthropic's average across real workloads (anthropic, oct 7 2026). that is the number to budget with, because real usage mixes both tiers.

there is a fourth 90% in the announcement, and it is not a price. anthropic says prompts up to 100,000 tokens made up around 90% of requests to the previous haiku model (anthropic, oct 7 2026). so most haiku traffic stays under the line, and the model is priced to keep it there. the comparison a business actually cares about, haiku 5.5 against sonnet 5.5, is a twentieth of the per-token price under the line ($0.10 and $0.50 against $2 and $10) and a quarter of it above the line ($0.50 and $2.50 against $2 and $10). i wrote about why per-token and per-task savings are different animals in you are using one ai model for everything.

where the 100,000-token line sits, and who crosses it

the line is measured per request, on the whole prompt, and a request over it pays the higher price on all of its tokens. anthropic's pricing page spells it out: a request's prompt length counts all of its input tokens, including cache reads and cache writes, a request over the threshold pays the higher prices even when part of its prompt is a cache hit, and earlier requests keep the prices they were billed at (anthropic pricing page, accessed oct 13 2026).

in plain terms: a long conversation re-sends everything that came before it as the prompt of every new turn, and so does a big cached system prompt. one short question at the end of a long thread with three attached pdfs is a long prompt. a weekly job that summarizes a folder of documents is a long prompt. neither feels expensive while you are doing it, and both get billed at the over-the-line rate once they pass 100,000 tokens.

two things soften this. on a claude.ai subscription you do not see a per-token bill at all, so the line only matters when the work is billed through the api, for example through an automation, a tool someone built for you, or a developer's integration. and for bulk jobs that can wait, anthropic's batch processing halves the price on both tiers, to $0.05 input and $0.25 output per million tokens under the line (anthropic pricing page, accessed oct 13 2026). classifying five thousand support tickets overnight is exactly that job.

where haiku 5.5 is not the cheap option

on complex multi-step work, haiku 5.5 scores a little over half of sonnet 5.5, and a model you have to run twice is not the cheap model. on terminal-bench 4.0, anthropic's chosen example of complex agentic coding, haiku 5.5 scores 39.2% against sonnet 5.5's 70.6% (anthropic, oct 7 2026). anthropic says that chart illustrates the important difference between haiku and its larger models, and that sonnet 5.5 and opus 5.5 remain the better choice for that kind of task (anthropic, oct 7 2026).

the cost of a wrong answer never appears on a price sheet. it appears as the rerun, the twenty minutes you spend fixing the output by hand, or the client who noticed before you did. if you cannot check the output fast, the cheap model costs more per finished task. i keep a short routine for that in how to verify ai output fast. the early customer reports in anthropic's announcement describe the haiku lane well. box's vp of ai products reported that in early testing, claude haiku 5.5:

"scored 11 points higher than haiku 4.5 at about half the latency"
yashodha bhavnani, vp of ai products, box, in anthropic's claude haiku 5.5 announcement, oct 7 2026

the same quote says where box would put it: analytical work that runs at scale, from cost reports to financial summaries and weekly recurring reviews (anthropic, oct 7 2026). recurring, bounded, lots of volume. that is a description of the haiku lane that a small business can use directly.

the four-step routing test

run this on your own work in about twenty minutes, and you will know which jobs move to haiku 5.5 and which stay on sonnet 5.5.

1. is it the same job every time? recurring, templated, bounded work (summaries, classification, lookups, the weekly report) is the haiku lane. a one-off job that changes shape halfway through is not.
2. will the prompt stay under 100,000 tokens? count the thread, the attachments, and the cached context, not just the question. if it crosses, trim the prompt or send it to sonnet 5.5, because haiku above the line costs five times haiku below it.
3. what does a wrong answer cost, and who catches it? if a subtle mistake would cost money, a client, or a legal problem and nobody checks the output, it stays on sonnet 5.5. anthropic's own line is that sonnet and opus remain better for complex multi-step work.
4. run ten real examples on both and compare cost per finished task. your work, not a benchmark. count the reruns and the fixes as part of the cost. per token, haiku wins by twenty to one. per finished task, you have to measure.

if you route across models automatically, read you might not be running the model you benchmarked before you trust the routing, because a fallback that silently swaps haiku for sonnet, or the other way round, breaks both the budget and the test.

the takeaway

haiku 5.5 makes a whole class of small-business work cheap enough to automate: the repetitive, high-volume, low-stakes jobs that were not worth running on a mid-tier model. it does not make sonnet 5.5 redundant. sonnet stays the default for long, ambiguous, and high-stakes work, and it stays the safer choice for any prompt that will pass 100,000 tokens. the mistake is not picking the wrong model. it is reading the headline price, moving everything, and finding out on the invoice which requests crossed the line.

newer to this and want the version with people in the room? the ai builders lounge is where the weekly builds happen.

or just follow along. new field notes most weeks on x, instagram, and tiktok.

browse all 30 free tools →