That's interesting. I wasn't aware that results can differ. Could you submit an issue here?https://github.com/Impactstory/unpaywall

I found an explanation on the oaDOI.org FAQ website. It says:

Why does the Unpaywall extension sometimes give different results from oaDOI?
Our Unpaywall browser extension uses oaDOI to find fulltext whenever you run into paywalled articles. It supplements oaDOI with other data sources, too; for instance, Unpaywall tries to parse and understand scholarly article pages as you view them. Consequently, Unpaywall's results are a bit more comprehensive than what you'd get by calling oaDOI directly.

https://oadoi.org/faq

This is still proving to be a troublesome issue for me, I'm interested in a CUP journal called 'Oryx' but roadoi results don't always match reality or that of what the Unpaywall browser plugin indicates

e.g. This article has a 'green padlock' with the browser plugin but is_oa = FALSE with roadoi / oaDOI API.

example <- roadoi::oadoi_fetch(dois = c("http://doi.org/10.1017/s0030605309990706"), email = "redacted@gmail.com")
> example
# A tibble: 1 x 16
                        doi best_oa_location     oa_locations data_standard is_oa
                      <chr>           <list>           <list>         <int> <lgl>
1 10.1017/s0030605309990706 <tibble [0 x 0]> <tibble [0 x 0]>             2 FALSE
# ... with 11 more variables: genre <chr>, journal_is_oa <lgl>,
#   journal_is_in_doaj <lgl>, journal_issns <chr>, journal_name <chr>,
#   publisher <chr>, title <chr>, year <chr>, updated <chr>, non_compliant <list>,
#   authors <list>

Can confirm that the API does not return an open access version for this DOI.

https://api.unpaywall.org/v2/10.1017/s0030605309990706?email=YOUR_EMAIL

According the Unpaywall FAQ, the extension does a live check.

The extension occasionally reports "open" when the API doesn't because the extension, in addition to checking our API, also checks the source code of publisher pages themselves as you visit them.
http://unpaywall.org/faq

Thanks, I've also reported it over at the impactstory repo ourresearch/oadoi#104