In my recent article on the price and performance of some commercial OCR (optical character recognition) models, I predicted Google’s Gemini 3.8 would cost $35.30 ± $23.70 to convert 1,000 pages containing computer printouts to text. Now, after scanning 5,877 pages, I can adjust this prediction to $34.02 ± $30.82, demonstrating the sample was representative. Details below.
In my evaluation, I found Gemini 3.7 and 3.8 to have the highest quality output (i.e. lowest CERiws and WER) of the admittedly limited list of models I tried. Since I ran that evaluation, Google has sunset Gemini 3.7, so I used 3.8 for my research task even though 3.8 had far worse cost efficiency.
Google’s published pricing as of this article’s date, assuming a standard tier API call on or after January 1, 2027 is:
Note that there are ways to reduce costs, such as making batch calls (50% reduction) and accepting higher latency.
I scanned 5,877 pages from the Community Memory corpus. The 28 documents consisted largely of scanned print-outs of messages posted to the digital bulletin board in the 1980s. Each call used the same prompt and included a PNG image of the page (extracted from the pdf using pdftoppm).
Each response included token usage data, reported as promptTokenCount (the input, including the tokenized image), candidatesTokenCount (the plain-text output), and thoughtsTokenCount (internal “thought process” tokens). The latter two types of token are priced the same. Google has some high-level description of the types of tokens and how to measure them, including how to calculate input tokens before any calls.
Plotting a probability histogram for each response yields the figure below:
Numerically, the token consumption per page:
| Token Type | Average (Mean) | Standard Deviation |
|---|---|---|
| prompt | 1277 | 10 |
| candidates | 774 | 194 |
| thoughts | 3507 | 3914 |
As expected, there is little variability in input tokens (prompt) as each request was very similar (pages were uniform in size, almost every page came from the same printer and used the same font). Similarly, the physical limits of a page limits the variability in output tokens (candidates), although plain text as the output format has higher variability than the image input. Thought tokens have extremely high variability, which is the same behavior we noted in our previous evaluation.
The three tokens are not correlated with each other. Graphically, they are not great fits of normal distributions, but do have bell shapes.
OCR prices are usually expressed in 1,000 pages. Scaling the above numbers to 1,000 pages and expressing two standard deviations (95.4%) of range, we get $34.02 ± $30.82.
| Token Type | Average per 1k pages | 2 Standard Deviations |
|---|---|---|
| prompt | $1.92 | $0.02 |
| candidates | $5.80 | $1.45 |
| thoughts | $26.30 | $29.35 |
| Total | $34.02 | $30.82 |
I’m unclear why Gemini 3.8 has such high variability in thought tokens compared to 3.7. I didn’t set thought level in the requests, so per the documentation, it should have defaulted to “medium”.
Thinking apparently has some impact on quality, at least anecdotally. During one of the sessions at the AI and History Conference, the presenter noted the latest models improved the quality of numerical table extraction by using printed column totals in the original to fix internal rows. It would be useful to benchmark OCR tasks by thinking level; variability might be decreased by setting the thinking level to “low” while quality maintained.
A potential research effort would be to seek some correlation between features of the input and thought tokens consumed. Perhaps less probable n-grams (e.g. spam-ish writing) might incur more cost? Could it be related to safeguards and similar systems triggered by the document content discussing something sensitive or criminal? However, there’s a damper on this form of research in that 1) Gemini models are a black-box, 2) Google has a habit of sunsetting models with little notice, potentially killing a research project mid-stream, and 3) difficulty in generalizing knowledge from a specific model to the field.