Claude Haiku 5.5 pricing changes the default. The old question was which tasks justified a model call at all. The new one is which tasks justify anything bigger than the small model. Start every classification, summary and triage job on Haiku, and make a task earn its way up to Sonnet or Opus with evidence from your own data.
What Claude Haiku 5.5 pricing changes about task selection
Anthropic says Haiku 5.5 costs around 75% less to run than Haiku 4.5 on average, and that it is built for high-volume, cost-sensitive work. Other coverage used a bigger number, for example Quartz’s “up to 90%” headline (8 October 2026). I would plan around the vendor’s own average, since “up to” describes the best case and your workload mix decides where you land.
I do not have the full price table in front of me, so I will not quote per-token rates here. Pull them from the pricing page and multiply by your real prompt sizes. Anthropic notes that prompts up to 100,000 tokens made up around 90% of requests to the previous Haiku, so most business jobs sit inside the range where the small model is priced best (same announcement).
The practical effect is on the borderline list. Tasks that were too thin to automate at the old price, such as tagging every inbound email or summarising every call, now clear the bar. The savings also change how often you can afford to run a job. A nightly pass can become an hourly one.
Tasks that move to the small model
Anthropic names summaries, compactions, database queries and classification as the core workloads. That matches what the launch customers describe. Box reported that Haiku 5.5 scored 11 points higher than Haiku 4.5 at about half the latency. AlphaSense, which runs about 8M calls a week on a document question-answering feature, measured 0.84 against 0.76 on 400 queries. HubSpot reported its best score yet on a CRM task suite, including a stale-record audit. All three are in the launch post, and all three are vendor-selected quotes, so read them as direction rather than proof.
For a small business, the candidates look like this:
- Classification at scale: routing inbound mail, tagging support requests, labelling leads by intent.
- Summaries that feed a person, such as call notes and weekly reports.
- Extraction from documents into fixed fields.
- First-pass ticket triage automation, where a wrong answer gets caught downstream.
Triage carries one risk. A cheap model makes it tempting to score every lead and discard the low ones. That is the failure described in automating lead triage without losing the good ones: scoring buries deals that do not resemble your last ten. Cheaper calls make it easier to route by confidence and send the uncertain middle to a person, instead of forcing a yes or no.
Cost per task is the number to compare
Compare cost per task, which a per-token rate does not give you. It includes retries, escalations to a larger model, and the human minutes spent fixing wrong output. A cheap model that needs a second pass on a fifth of its jobs can lose its advantage quickly. I do not know your retry rate, and neither does anyone else until you measure it.
So run a small evaluation before moving anything. Take a few hundred real examples with known correct answers, run them on the small model and the larger one, and record accuracy, retries and review time. The method in measuring quality, not just speed applies directly. If the small model matches within a margin you can live with, switch. If not, you have a number to point at rather than a feeling.
Haiku 5.5 is also the first Haiku-class model with an adjustable effort setting, according to the announcement. That adds a second dial. Before you escalate a failing task to a bigger model, try a higher effort setting on the small one and see whether the failures go away.
What stays on the larger model
Keep ambiguous judgement, long multi-step planning and anything where one wrong answer is expensive on the larger models. Anthropic itself pairs Haiku with Opus 5.5 and Sonnet 5.5 as a subagent on coding work. That is the intended split: the big model plans, the small one does the repetitive legwork. Rogo describes the same pattern, with a larger model building a deck while a Haiku subagent pulls a revenue line from a 10-K.
Not every task needs a rewrite to benefit from the new tier. Anthropic also halved Sonnet 5.5 cache read prices, which it says makes Sonnet around 20% cheaper on most agentic work (source). If your work needs the mid-tier model and reuses a long prompt, caching may save you more than switching down.
Low-volume work is a poor fit for any of this. If a task runs fifty times a month, the price difference is noise and the engineering time to route it costs more than it saves. Leave it on whatever model works best.
Build model routing with a model downgrade ladder
Model routing is the practical form of all this: a default model, a rule for when to escalate, and a record of why. Start at the bottom of the ladder and climb only on a failed check, such as a schema validation error or a low confidence flag. The model downgrade ladder in keeping token spend predictable covers the budgets and hard caps that stop token spend from drifting as volume grows.
Small models are less forgiving of loose instructions, so prompts matter more at this tier. Explicit output formats and named failure modes do most of the work, as laid out in prompting for reliability, not cleverness. Put the model name in configuration, not in code. Prices and rankings are moving fast. Inception claimed on 20 October that its Mercury 3 beats both Haiku 5.5 and GPT-6 Luna at much higher speed (SiliconANGLE (SiliconANGLE, October 2026)), and I have not tested that claim. Independent head-to-heads exist too, such as Buoy’s 256-job comparison against GPT-6 Luna, and they are worth reading with the scoring method in mind before you trust a winner.
Where I would be careful
Vendor benchmarks and customer quotes describe the vendor’s chosen tasks. Your documents, your tone and your edge cases are different, which is why the evaluation above comes first. Switching a model on a regulated or customer-facing decision without a human check is a separate risk that no price cut changes.
The summary is narrower than the headlines. The floor price for useful automation dropped, so more routine work is worth doing. Whether a particular job of yours is one of them is an empirical question, and you can answer it in an afternoon with a few hundred labelled examples.