Cohere Parse 5: Losing the Benchmark, Winning on Cost Per Page
Cohere released Parse 5 this week: a 2.3B-parameter vision language model that does one job — turn PDFs, slides and images into structured Markdown. Tables come back as HTML, forms as key-value pairs, images with descriptions and bounding boxes, everything in reading order. Flat pricing at $1.50 per 1,000 pages, processing 4.5 pages a second, available via API and Microsoft Foundry.
VentureBeat's framing was unusually honest for a launch cycle: Parse 5 loses the benchmark on points and wins on cost per page. Cohere isn't claiming the most accurate parser, they're claiming the right cost-capability mix at enterprise scale — a 2.3B model small enough to run cheap, good enough that the errors don't matter at $1.50 per thousand pages. After a year of every lab claiming SOTA on everything, positioning a launch on 'good enough, much cheaper' is almost refreshing.
Why it matters for agents: document parsing is the mouth of every enterprise agent. Agents don't fail on reasoning nearly as often as they fail on a table that got mangled into a character soup somewhere upstream. The parsing layer has quietly become competitive infrastructure — and pricing it like a commodity utility, rather than an AI product, is probably what wins that layer.
Details: https://cohere.com/blog/parse
← Back to all articles
VentureBeat's framing was unusually honest for a launch cycle: Parse 5 loses the benchmark on points and wins on cost per page. Cohere isn't claiming the most accurate parser, they're claiming the right cost-capability mix at enterprise scale — a 2.3B model small enough to run cheap, good enough that the errors don't matter at $1.50 per thousand pages. After a year of every lab claiming SOTA on everything, positioning a launch on 'good enough, much cheaper' is almost refreshing.
Why it matters for agents: document parsing is the mouth of every enterprise agent. Agents don't fail on reasoning nearly as often as they fail on a table that got mangled into a character soup somewhere upstream. The parsing layer has quietly become competitive infrastructure — and pricing it like a commodity utility, rather than an AI product, is probably what wins that layer.
Details: https://cohere.com/blog/parse
Comments