Sonnet 5 vs Haiku 4.5: when the cheaper Claude is enough

By LW Forge — maintainer of LLM Scout · Updated August 10, 2026

Opus is a niche choice for most teams. The decision that actually shows up on the invoice, month after month, is Sonnet 5 versus Haiku 4.5 — and because Anthropic prices both tiers with the same input-to-output ratio, Sonnet costs exactly three times Haiku on identical traffic. No modelling required: whatever your bill is on one, multiply or divide by three for the other.

Sonnet 5 is running a launch promotion of $2/$10 per million tokens through August 31, 2026, which narrows the gap to 2× while it lasts. Plan on the list rates above — the promo expires, the workload doesn't. Both tiers share the same 200k-token context window, so moving down to Haiku never costs you room for a long document; you can confirm what fits in the context window calculator.

Where the 3× actually bites

Consider a document enrichment pipeline: one million documents a month, each sending about 800 input tokens and generating about 60 output tokens of structured labels. That's 800M input and 60M output tokens:

$2,200 a month, for a task where the output is a handful of labels drawn from a fixed vocabulary. Add the Batch API — roughly half price when results can wait — and Haiku lands near $550 against Sonnet's $1,650. That's the shape of workload where the cheap tier isn't a compromise: the task is narrow, the answer space is small, and correctness is measurable rather than a matter of taste. Price your own volumes in the Claude cost calculator.

The gap matters far less in the opposite shape. A product feature serving 20,000 requests a month at the same token sizes costs $66 on Sonnet and $22 on Haiku. Saving $44 by risking the quality of something users read is a bad trade — at low volume, pick the better model and spend your attention elsewhere.

Tasks where Haiku genuinely matches

Haiku holds up well where the job is bounded and the correct output is checkable: classification into a known set of categories, routing and triage, extraction into a fixed schema, sentiment and language detection, deduplication, short factual summaries of a single source, moderation and spam filtering. What these share is that you can build an eval set of a few hundred labelled examples and measure accuracy directly, instead of arguing about whether the prose feels better.

Sonnet pulls ahead where the task is open-ended: multi-turn conversation with real users, drafting that has to hold a voice, code changes across several files, chains of tool calls where a single misread compounds, summarizing conflicting sources, anything requiring the model to notice that the question itself is ambiguous. The tell is whether a wrong answer is obvious and cheap to catch, or subtle and expensive to miss.

Test it before you assume it

The honest way to settle this is boring and takes an afternoon. Pull 200-500 real requests from production. Run them through both tiers. Score the outputs against whatever "correct" means for your task — exact match for classification, human review for anything softer. If Haiku matches Sonnet within your tolerance, the 3× is pure margin. If it doesn't, you now know exactly where it breaks, which is more useful than a general opinion about small models.

Two details that skew this test if you skip them: prompts tuned for a larger model often need more explicit instructions and a couple of examples to work on a smaller one, so re-tune before concluding Haiku can't do it. And measure the output length — smaller models sometimes compensate with longer, more verbose replies, and output is billed at 5× input on both tiers. Check real token counts in the token calculator rather than trusting a word estimate; why 1,000 words are never 1,000 tokens covers why the two diverge.

The routing pattern that captures most of the savings

You rarely have to choose globally. Run Haiku first, have it emit a confidence signal or a self-check, and escalate only the requests that fail it. On the pipeline above, escalating 10% of traffic to Sonnet costs about $1,320 a month against $3,300 for all-Sonnet — roughly 60% saved, with the hard cases still handled by the stronger model. The overhead is one extra branch in your code and a threshold you tune with real data.

Stack the other levers on top: cache the stable prefix of your prompt at about a tenth of the input rate (prompt caching), and batch anything that isn't interactive for roughly half price. Applied together, these usually beat what you'd save by agonizing over the tier in the first place. If you're weighing all three tiers rather than this pair, Opus vs Sonnet vs Haiku has the full table, and what a chatbot really costs works through an end-to-end example.