Two cents per thousand units.
Then the alternatives, priced as a stack.

A rate per minute, per page or per character only becomes a rate per hour, per thousand or per gigabyte once somebody says how. Here is how.

Audio: batch transcription of one channel, with no diarisation, redaction or custom vocabulary; streaming is dearer at every vendor; Google sits between the two at about $0.96. Video: one pretrained feature over the video - label detection - after the first 1,000 minutes a month; a second feature is charged again at the same rate, and Data Oil's figure is one frame a second, so a finer sampling costs proportionally more. Documents: pay as you go with no volume commitment, and no OCR add-on for a scan; Azure and Google list the same rate, and both discount a committed tier. Translation: one byte to the character, so a gibibyte of text is about 1,074 million characters, and the charge falls again on every target language. Text in a multi-byte script - CJK, Cyrillic, Greek, Devanagari - holds fewer characters to the gigabyte and costs proportionally less at these vendors, while Data Oil meters the bytes either way. Vendor list prices read on 16 September 2026; the pages are in the sources below.

Every figure is a month. A search is a vector or full-text query; reads and writes are statements; questions are put to the assistant. The workloads are representative shapes chosen for this comparison. The column headings were four words each until the table stopped fitting on a laptop, which is a worse way to lose a reader than a caption is.

A quarter of a million records, two million reads, half a million searches a month.

Supabase Pro is cheaper on this workload, at $40.34 against our $123.80 - and at this size it should be. A flat monthly fee wins until the workload outgrows the compute it includes. What it does not include is ranked search, a graph, translation, media or an assistant: the day any one of those is wanted, it becomes a second product beside the first.

Two products are priced here but kept off the table: Firestore Standard at $2.97 (no full-text search at all - Google's own guidance is to add a third-party engine); Cloudflare D1 + Workers Paid at $9.30 (FTS5 inside, but vectors are a separate product). Both come out below every row above, and neither can run this workload as written - so putting them in the same column would be a cheaper number for a different job.

Fifty thousand documents, a hundred thousand searches, and twenty thousand questions answered with citations.

Two products are priced here but kept off the table: Firestore Standard at $7.02 (no full-text search at all - Google's own guidance is to add a third-party engine); Cloudflare D1 + Workers Paid at $33.46 (FTS5 inside, but vectors are a separate product). Both come out below every row above, and neither can run this workload as written - so putting them in the same column would be a cheaper number for a different job.

The assistant is in none of these columns, ours included. Answering 20,000 questions a month with citations adds $192.00 to the Data Oil bill, at $0.005 a question against your own model. It adds nothing to the others because none of them answers a question - you would be buying a fourth product and wiring it to the first three. We have left it out of every column rather than set our whole bill against their partial one.

Five million records, five million reads and a million writes a month, read by BI tools over the PostgreSQL port.

Supabase Pro is cheaper on this workload, at $28.03 against our $135.00 - and at this size it should be. A flat monthly fee wins until the workload outgrows the compute it includes. What it does not include is ranked search, a graph, translation, media or an assistant: the day any one of those is wanted, it becomes a second product beside the first.

Two products are priced here but kept off the table: Firestore Standard at $11.43 (no full-text search at all - Google's own guidance is to add a third-party engine); Cloudflare D1 + Workers Paid at $25.41 (FTS5 inside, but vectors are a separate product). Both come out below every row above, and neither can run this workload as written - so putting them in the same column would be a cheaper number for a different job.

A stack price is a model, and a model that does not say what it assumed cannot be checked. Five choices decide every figure in the three tables above, and each of them is on this page because most of them make the alternatives cheaper rather than dearer.

Each row of each table also carries the plan, the allowance and the sizing behind its own figure. The vendors and the pages they were read from are in the sources below.