Breeze credits or your own API key: what AI actually costs at 150 people

Short answer: a Breeze action costs 10 HubSpot Credits, which is 10 cents. The same job through a direct API call costs between a third of a cent and two cents. That is a 10x to 30x difference per task. It still does not mean you should move, because the direct call needs an engineer and the credits do not, and on the assumptions below the crossover does not arrive until around 140 seats.

That is the answer. The rest is the working, because nobody defending this budget should have to take our word for it.

Where these numbers come from. Every figure below is built from published rates, not from a private benchmark: HubSpot's own Credits rate sheet, Anthropic's pricing, and OpenAI's. The volume assumptions are stated in full so you can swap in your own. Redo it with your numbers and you should land where we did.

We keep getting asked this in the same room, usually by an ops lead at a 120 person business who has been handed an AI line item for next financial year and no way to defend it. HubSpot quotes credits. Anthropic and OpenAI quote tokens. The two numbers are not comparable, and no vendor has any interest in making them comparable.

The three jobs we priced

We picked jobs a revenue team actually runs every day, not benchmarks:

• Summarise a sales call from a transcript into five bullet points and a next action.
• Draft the follow-up email from that summary, in the rep's voice, with the deal context attached.
• Enrich a company record from a website and a LinkedIn page into six structured fields.

On the Breeze side all three are the same line item: executing one Breeze action in a workflow costs 10 HubSpot Credits, and a credit is one cent. So every job below costs ten cents, whether it is summarising six thousand words or filling in six fields. On the API side you pay for what you actually send.

The API column is Claude Haiku 4.5 at one dollar per million input tokens and five per million output. Move to a mid-tier model and the per-run cost roughly triples, which still leaves it three to eleven times cheaper than the credit. Prices are USD.

The assumptions, so you can argue with them

• Each seat runs 2 call summaries and 4 follow-up drafts a day, over 21 working days.
• Record enrichment runs at 500 records a month for the whole portal, not per seat.
• A 30 minute call transcript is about 4,500 words, which is roughly 6,000 tokens.
• Your HubSpot tier includes some credits each month. Professional includes 3,000. That is 300 Breeze actions, which one seat gets through in about ten weeks.

At 20 seats that is 3,020 actions a month. Credits: 30,200 consumed, 3,000 included, so $272 in overage. Direct API: $13. You save $259 a month.

At 150 seats it is 19,400 actions. Credits: 194,000 consumed, $1,910 in overage. Direct API: $86. You save about $1,824 a month, or roughly $22,000 a year.

Why $259 a month is not a reason to move

Because owning the direct call is not free. Somebody writes it, monitors the spend, and fixes it when a model update changes the output. Call that a five day build and a day a month of maintenance, and at a realistic loaded rate you are carrying something like $1,700 a month once you amortise the build over a year.

Against a saving of $12 per seat per month, that pays for itself at about 140 seats. Below that, credits are the cheaper answer and it is not close. This is the number the article exists to give you, and it is the one most people guess wrong in both directions.

Why the per-run number is not the decision

Here is the part the token comparison hides. A direct API call has a build cost and an owner. Somebody writes the prompt, handles the failures, watches the spend, and fixes it when the model behaves differently after an update. That is real money and it does not show up in a price per million tokens.

Breeze has none of that. It is already wired into your records, it respects your permissions, and when HubSpot changes something they fix it. You are paying a premium for somebody else to own the thing. For a lot of work that is a good trade and we will say so to a client's face.

That $1,700 is the real variable. It falls if the job is simple and rises if it is not, which is why the premium stops being worth it only when three things are true at once:

1. The job is identical every time. Same input shape, same output shape, no judgement. That is where a tuned direct call pulls ahead, because you can strip the prompt down and cache the parts that never change.
2. The volume is high and predictable. Thousands of runs a month, not dozens. Below that, engineering time swamps the saving.
3. The output feeds a system, not a person. If a human reads every result, you are paying for convenience anyway and Breeze is the convenience.

The costs nobody quotes you

Whichever way you go, four things move the number and none of them appear on a pricing page.

Retries. Every failed or malformed response is paid for. Budget for a failure rate, because the honest one is not zero.

Context you did not mean to send. The single biggest cost blowout we see is a prompt that quietly attaches an entire deal history when it needed the last three notes. Same output, several times the spend.

The pilot that never turns off. Somebody builds a proof of concept, it works, it stays running against real records for eight months and nobody owns the bill.

Model changes. Prices fall, but capability moves too. Whatever you cost today is a snapshot. Re-check quarterly and write the review into somebody's calendar or it will not happen.

What we would actually do at 150 people

Run both. Not as a hedge, as an architecture.

Keep Breeze for everything a human reads and everything that touches a record inside HubSpot, because the integration is the product and rebuilding it yourself is a bad use of an engineer. Move the high volume, machine-to-machine work to a direct call: bulk enrichment, classification, anything that runs overnight and produces a field rather than a paragraph.

Then measure it. Not once. The line between the two moves every time either vendor changes a price, and it has moved twice this year already.

The honest version: most businesses asking this question do not have a cost problem yet. They have an unmeasured problem. If you cannot say what your AI spend bought you last month, the answer is not a cheaper model, it is a meter.

Before you sign anything

Three questions, in this order. What is the job, exactly, and is it the same job every time? How many times a month, honestly, not aspirationally? And who owns it in six months when the person who built it is on leave?

If you can answer all three, the cost model picks itself. If you cannot, no pricing page will save you.