The bottleneck Cohere names is narrow, and it is real. Kyle Lastovica, a Cohere spokesperson, told The Deep View that standard optical character recognition misses tables, diagrams and visual structure, and that many companies patch the gap by pointing frontier models at the same pages, which drives spend sharply upward [8]. He also frames input quality as the thing that matters most now that agents and retrieval-augmented generation sit downstream [9]. That is the shape of most ingestion pipelines I have worked on: a cheap deterministic pass, then something expensive to repair what the cheap pass flattened. Parse is sold as the tier in between [3].
Both published rates are denominated in pages, so a year of extraction can be estimated from a document count that is already sitting in a database. The ratio works out to 15 percent [14], and the delta is $8.50 per 1,000 pages [15].
The number that decides the design is the escalation rate. Assume you keep a fallback to the expensive tier for pages the cheap model cannot finish. Blended cost per 1,000 pages is then 1.50 + 8.50f, where f is the escalated share. At f = 0.2 you pay $3.20 [16]. At f = 0.41 you are paying half the frontier rate [17] while also maintaining a second code path and whatever logic decides when to take it. That number, f, is a property of your own documents, and no vendor's spec sheet can tell you what it is.
Cohere's price-performance ranking is the seller grading its competitors on a corpus the seller chose [6]. For it to transfer, your pages have to fail where its pages fail, and your language mix has to fit the nine it ships [7]. The generous reading is the one I would actually use: at this price, running your own eval is cheap enough to be the first task rather than the deferred one.
The routing half of the same bet is thinner in the reporting. The Deep View does not say how Replit's router decides which model gets a request, so there is nothing to price yet. It reads NextLM's work on purpose-built, efficient models for sales tasks as the same thesis [12], and says months of token bills have left some buyers wary of bulky general-purpose models [13]. Document parsing is not the flashiest use case, as The Deep View concedes [18], and it is exactly the kind of high-volume work where per-page arithmetic decides the bill.
In my context the sequence is: pull last quarter's page counts, label a sample large enough to estimate f, then price both tiers against that sample. Checking a vendor's per-page rate takes an afternoon, but estimating f takes closer to a week.
