Haiku, Sonnet or Opus: which Claude for which HubSpot job?
Anthropic released Claude Haiku 5.5 on 7 October and cut Sonnet's cache price the same day. Here is which model we put on which HubSpot job, and what each run costs in Australian dollars.
Use Haiku 5.5 for the jobs you run hundreds of times a day: sorting leads, cleaning records, summarising calls and drafting customer replies. Use Sonnet 5.5 for weekly work that needs judgement across a few dozen records, like a pipeline review or account research. Save Opus 5.5 for the long, hard jobs where a wrong answer costs real money: win/loss analysis, tenders and writing code that runs inside HubSpot.
The price gap between them is large. Per million tokens, on Anthropic's price list, Haiku 5.5 costs US$0.10 to read and US$0.50 to write. Sonnet 5.5 is 20 times that, and Opus 5.5 is 40 times that. Put a high-volume job on the wrong model and you pay 40 times more than you need to. Put a hard job on the wrong model and you pay for it in rework.
Every Australian dollar figure below uses the RBA rate of 0.6979 US dollars per Australian dollar from 9 October 2026, with nothing added for tax. Anthropic lists API prices in US dollars, excluding any tax.
Output price per million tokens, USD, excluding tax. The Haiku price is for prompts up to 100,000 tokens. Source: Anthropic pricing. Prices checked 12 October 2026.
What changed this week?
Three things, all on 7 October 2026, according to Anthropic's release notes.
Haiku 5.5 launched. It is Anthropic's model for high-volume, fast work. It reads up to a million tokens at once, can write up to 128,000, and is the first Haiku that lets you turn its effort up or down. Its price is US$0.10 per million input tokens and US$0.50 per million output tokens, for prompts up to 100,000 tokens. Over that, both prices go up five times.
Sonnet 5.5 cache reads halved, from US$0.20 to US$0.10 per million tokens. More on why that matters below.
Team and Max plans now come with monthly API credits. On Team, each Standard seat adds US$20 and each Premium seat adds US$100, pooled across the team and capped at US$500 a month (about A$716). The credits do not roll over, and Enterprise plans are not eligible.
Anthropic's Haiku page also quotes HubSpot's engineers saying it got the best score they have seen on their CRM test suite. That is a launch-page claim, so test it on your own records.
Which Claude model should do which HubSpot job?
Here is where we start for a 60 to 200 person business. The cost column is a worked estimate, with the token counts and the maths shown under the table. Your prompts will be shorter or longer, so treat it as an order of magnitude.
Prices checked 12 October 2026.
| HubSpot job | Model | Why | Rough cost per run |
|---|---|---|---|
| Lead triage: is this enquiry a buyer, and which service? | Haiku 5.5 | Short text, a fixed set of answers, needs to be fast | A$0.0003 (about A$0.32 per 1,000 leads) |
| Data clean-up and enrichment on a company record | Haiku 5.5 | Thousands of records, simple rules, can run overnight | A$0.0006 (A$6.45 per 10,000 records, half that in a batch) |
| Call transcript to CRM note | Haiku 5.5 | Summarising one document is what it was built for | A$0.003 |
| Customer agent reply from your help articles | Haiku 5.5 | Speed matters, and the knowledge base can be cached | A$0.001 with caching |
| Weekly pipeline review | Sonnet 5.5 | Judgement across 50 to 100 deals, once a week | A$0.16 |
| Target account research with web search | Sonnet 5.5 | Several searches, then a sensible summary | A$0.36 |
| Win/loss analysis across a quarter of closed deals | Opus 5.5 | Long reading, patterns across many deals, the answer changes strategy | A$1.00 |
| Tender response draft | Opus 5.5 | Long documents in, long careful writing out | A$1.03 per full pass |
| Writing a HubSpot custom code action | Opus 5.5 in Claude Code | Code that runs on your live data has to be right | about A$0.99 per instruction |
Worked estimate, using the token counts listed under the table and the RBA rate on 9 October 2026, with nothing added for tax. The Haiku jobs barely show on this scale, and that is the point. Prices checked 12 October 2026.
How the estimates were worked out. Cost is input tokens times the input price, plus output tokens times the output price.
- Lead triage: 1,500 tokens in, 150 out on Haiku. (1,500 × $0.10 + 150 × $0.50) ÷ 1,000,000 = US$0.000225.
- Company clean-up: 3,000 in, 300 out on Haiku = US$0.00045.
- Call note: a 45 minute call is roughly 15,000 tokens of transcript, with 800 tokens of note out. On Haiku that is US$0.0019. The same job on Sonnet is US$0.038, twenty times more.
- Customer reply: 20,000 tokens of help articles read from cache at US$0.01 per million, plus 5,000 fresh tokens and 400 out = US$0.0009. Without the cache it is US$0.0027.
- Pipeline review: 40,000 tokens of deal export in, 3,000 out on Sonnet = US$0.11.
- Account research: 60,000 in, 3,000 out on Sonnet, plus ten web searches at US$10 per 1,000 = US$0.25.
- Win/loss: 150,000 in, 5,000 out on Opus = US$0.70.
- Tender: 80,000 in, 20,000 out on Opus = US$0.72.
- Custom code: this one is measured, not estimated. Our Opus 5.5 logs came to US$0.69 per instruction at list price.
Two things push real costs up. All three models now think before they answer by default, and thinking is billed as output. On a simple triage job, set Haiku's effort to low. Second, the newer models count text with a new tokenizer. Anthropic says it produces about 30% more tokens for the same text. Measure your own prompts before you budget.
Is Haiku good enough for lead sorting?
For most teams, yes. One caveat comes from our own testing.
When we ran 184 real enquiries through Jev, a decision model that only answers questions and never writes, it sorted them in 9 seconds for under one US cent and got 92.7% right. At our worked estimate, Haiku 5.5 would cost about six Australian cents for the same 184. Both figures are tiny. The difference is what you get back. Jev gives you a probability, which tells you which leads a person should check. Haiku can also write the note to the rep and the first reply.
The lesson from that test applies to Haiku too. Four narrow questions beat one broad one. "Do they describe a need of their own?" gets a more reliable answer than "Is this a good lead?" whichever model you ask.
Not sure which model your HubSpot jobs need?
Send us your list of AI jobs. We will tell you which model each one belongs on and what it should cost you a month.
Most work stops at Haiku. Only the jobs that need more judgement or a long read move up a size.
When is Sonnet the right call?
Sonnet 5.5 sits in the middle on price and handles well-defined jobs that need some judgement. Anthropic's own guidance puts it this way: Opus 5.5 for complex coding and knowledge work, Sonnet 5.5 for well-scoped tasks.
A weekly pipeline review is the clearest example. The job is the same every week, the prompts can be written once, and the output is a short list a sales manager reads before the meeting. Our seven pipeline review prompts run well on Sonnet. At about 16 Australian cents a run, a year of weekly reviews costs less than one lunch.
Target account research is the other good fit. The model searches, reads and summarises, and Opus rarely finds anything Sonnet missed on a job like that.
When is Opus worth paying for?
When the job is long, the stakes are high and a person would otherwise spend a day on it.
Answering a tender means reading a hundred pages of requirements and writing answers that have to be accurate about your business. A dollar a pass is nothing next to the cost of a wrong compliance answer.
Writing custom code actions is the same story. That code changes records in your live portal. Opus 5.5 also made far fewer wrong turns in our logs: interruptions fell from 13.7 per 100 messages on Opus 5 to 1.5.
Win/loss analysis is a judgement call between Sonnet and Opus. We use Opus because the output changes what the sales team does next quarter, and the whole run costs about a dollar.
How do prompt caching and the Batch API cut the bill?
Prompt caching stores the start of a prompt so the next call can read it cheaply. A cache hit costs 10% of the normal input price on Haiku, and 5% on Opus 5.5 and on Sonnet 5.5, which dropped to that level this week. It pays off when many calls share the same long opening within a short window. The cache lasts five minutes by default or an hour if you pay a little more to write it.
That makes it a good fit for a customer agent, where every reply starts with the same help articles, and for a tender, where you ask twenty questions about the same document in one sitting. It does nothing for a weekly pipeline review, because the cache is long gone by next Tuesday.
The Batch API halves the price of input and output in exchange for waiting. You send a pile of requests and collect the answers later. Overnight enrichment, monthly data clean-up and a quarterly win/loss pass all suit it. Lead triage and customer replies do not, because a lead that waits hours for a reply goes cold. Our post on speed to lead has the numbers on that.
Will the new Team credits cover this?
For most of the jobs in the table, yes. A team on 20 Standard seats gets US$400 a month, about A$573. At our worked estimate that would triage well over a million leads on Haiku. Even a busy month of tenders, pipeline reviews and win/loss on Opus and Sonnet usually costs tens of dollars.
Two things to know before you plan around them. The credits expire each billing month, so they suit steady work and do not cover a one-off spike. And when they run out, API requests stop unless you turn on auto-reload. That is a useful spending cap for a pilot, and a problem for a live customer agent. Pick one on purpose.
For the bigger picture on what AI costs a business your size, including HubSpot's own Breeze credits, read Breeze credits or your own API key. For getting the whole team using Claude, see how to roll Claude out to a 100-person business.
Frequently asked questions
Is Claude Haiku 5.5 cheaper than Haiku 4.5? Per token, much cheaper: US$0.10 and US$0.50 per million against US$1 and US$5. The new tokenizer counts the same text as more tokens, which takes back some of that saving. It is still a large drop for prompts under 100,000 tokens.
What happens if a Haiku prompt goes over 100,000 tokens? The whole request is billed at the higher rate, US$0.50 in and US$2.50 out per million, even if part of it is cached. Keep Haiku jobs short, and send long documents to Sonnet or Opus.
Do these prices include GST? No. Anthropic lists API prices in US dollars, excluding any tax. Its help centre says tax is worked out from your billing address, and you can add a tax ID such as your ABN under Settings, then Organization, in the Claude Console. It does not spell out how Australian GST shows on an API invoice, so check your own invoice and ask your accountant how to treat it in your BAS. Every Australian figure in this post is a straight conversion at the RBA rate on 9 October 2026, with nothing added for tax.
Can the data stay in Australia? Yes, if you run Claude through Amazon Bedrock. As at 12 October 2026, the AWS Bedrock model card lists an Australian cross-region option for Haiku 5.5, Sonnet 5.5 and Opus 5.5 that keeps requests inside its Sydney and Melbourne regions. Anthropic's pricing page says regional endpoints on Bedrock cost 10% more than global ones. Our post on where your data goes when you use AI covers the rest.
Sources
- Pricing, Anthropic, for model, cache, batch and web search prices.
- Release notes, Anthropic, for the 7 October 2026 changes.
- API credits for subscribers, Anthropic, for Team plan credits.
- Claude Haiku, Anthropic, for the HubSpot test suite claim.
- Exchange rates, Reserve Bank of Australia, for the 9 October 2026 rate.
- Billing address and tax calculation, Anthropic help centre, for how tax is worked out.
- Claude Haiku 5.5 model card, AWS Bedrock, for Australian cross-region availability.
Prices and models change often, so we will update this post when they do. In the meantime, which HubSpot job does your team spend the most hours reading and clicking through? That is usually the one to hand to Haiku first.
Want help moving the right jobs onto the right model? See our AI engineering work, or talk to us.
What would you put on Haiku first?
Start with the job your team does a hundred times a week by reading and clicking.