Our Customers Now Get 3x More Indexed Papers, for the Same Price
We moved the pipeline that reads and indexes your PDFs to OpenAI's gpt-5.6-luna. Indexing costs about 60% less, runs in under half the time, and we measured it against our own evaluation loop before shipping to make sure nothing got worse.
We moved the pipeline that reads your PDFs and turns them into searchable, citation-linked passages onto OpenAI's new gpt-5.6-luna. The result: indexing costs about 60% less, finishes in under half the time, and every paid plan now covers more than 3x the papers it did last week, at the same price. No plan changed price. No feature was removed. We just measured the swap carefully before shipping it.
What actually changed
| Before | After | |
|---|---|---|
| Cost per page indexed | ~2 credits | under 1 credit |
| Time to index a typical paper | baseline | under half |
| Papers per month, Starter plan | ~15 | ~50 |
| Papers per month, Pro plan | ~45 | ~150 |
Those plan numbers are already live on our pricing page. We didn't raise prices to capture the savings, we passed them through.
How we knew it wouldn't get worse
A cheaper model is only a good trade if the output holds up, so we don't ship a model change on cost and speed numbers alone. Before and after the swap, we ran the same fixed set of real academic PDFs through the full pipeline and read the actual output: does the generated summary say only what the paper says, are the page citations pointing at the right pages, is anything invented that isn't in the source.
That last check caught something concrete. In the old pipeline, a paper's running header, its title repeated at the top of a references page, was occasionally mistaken for the start of a new section, and the model wrote a plausible-sounding summary for content that was actually just a bibliography. Under the new model, across every repeated test, that specific mistake didn't happen once. The references section got recognized and skipped cleanly instead.
We also caught a smaller bug in our own setup along the way: a newer prompt-caching option we tried turned out to silently disable caching entirely rather than guarantee it, something we only found by inspecting the raw API response directly instead of trusting the aggregate cost number. The version we shipped uses the caching mode that actually works.
None of this is a one-time check. We keep the same test set on hand specifically so that the next model or pipeline change gets checked against a real before-and-after, not a hunch that it's probably fine.
What this means for you
If you're on a paid plan, you already have more indexing capacity than you did last week. If you're indexing a large library through Zotero, the estimated cost you see before syncing now reflects the real number, not the old one. Nothing to configure. It's already running.
New posts, straight to your inbox
No newsletter fluff, just an email when we publish something new.
Email me when a new post is published on the Agent Bayes blog. You can unsubscribe anytime. We'll first send a confirmation email, and we only use your details for this. See our Privacy Policy.