A finance team handed us forty thousand scanned PDFs last spring and asked for clean Markdown. We priced the job against a frontier model, then quietly closed the spreadsheet. Cohere's Parse 5 is aimed at exactly that flinch: generally available since 27 August 2026, $1.50 per 1,000 pages, a 2.3B vision model that reads one page and returns Markdown with bounding boxes. It is not the best parser on the board and Cohere does not pretend otherwise. What we want to flag is the 79.2 ParseBench figure in the launch post. It averages three of the benchmark's five dimensions, and the two left out are the two where models of this shape usually fall over.
The short answer
Parse 5 turns a document page into Markdown with bounding boxes, cheaply and fast. The benchmark number Cohere leads with is a partial average, and the parts it skips are charts and visual grounding. Useful model, read the footnote before you promise anyone accuracy.
The number, and what is missing from it
Cohere leads with 79.2 on ParseBench and puts the component scores right there in the post: 87.0 on tables, 86.6 on content faithfulness, 64.0 on semantic formatting. Add those and divide by three and you get 79.2 exactly. That is how you know the other two dimensions are not in the average.
ParseBench has five. The two absent ones are charts and visual grounding, and they are not decorative. The benchmark’s own paper, arXiv 2604.08538 from the LlamaIndex team, is fairly blunt that grounding is where vision language models come apart: it reports GPT-5 Mini and Haiku below 8% on that dimension while older layout-detection parsers land somewhere in the 55 to 80 band. Grounding is the thing that lets an auditor trace an extracted figure back to a spot on a page. In regulated workflows that is not a nice-to-have.
Cohere does advertise bounding boxes as a feature, so the capability exists. It just is not in the number on the poster. Honestly, I do not read that as dishonest so much as normal launch behaviour, and Cohere published the breakdown that lets you catch it, which is more than most do.
Image: Cohere, from the Parse 5 announcement
The price is the actual product
$1.50 per 1,000 pages works out to $0.0015 a page. That is the whole pitch, and it is a good one, because document conversion is the least glamorous part of any retrieval pipeline and the one that quietly eats the budget.
Cohere’s own worked example is a company running 13 million pages a month. We checked the arithmetic and it holds together: 13 million pages at $0.0015 is $19,500 a month, so $234,000 a year on the API. Cohere says Model Vault saves that customer about $144,000 a year, which puts the single-tenant deployment near $90,000. Against hyperscaler document services at $10 per 1,000 pages, the same 156 million pages a year would run $1.56 million, and $1.56 million minus $90,000 is the $1.47 million saving Cohere quotes. The numbers are internally consistent, which is not always true of vendor cost slides.
The comparison against GPT-5.5 is a different animal. Cohere models a financial services workflow at 750 million documents a year and claims more than 98% cost reduction. Treat that as an estimate rather than a measurement, because it depends entirely on how many tokens a page turns into, and nobody publishes that number honestly.
What it is, mechanically
2.3 billion parameters, roughly 4.6 GB, built on Cohere Labs’ North-Micro-Vision-Instruct. Context window is 8,192 tokens. You hand it one page at a time as a base64-encoded data URI (our Base64 encoder is right here if you want to eyeball what your client is actually sending) and it returns Markdown: text in reading order, tables as HTML, lists, form key-value pairs, image descriptions, box coordinates.
That 8,192 figure matters more than it looks. This is a per-page tool. Whatever stitches pages back into a document, dedupes headers, or works out that a table runs across a page break, you are writing.
Throughput is 4.5 pages a second on one GPU, or 36 a second on an eight-card H100 node, which Cohere frames as 1.4x faster than dots.mocr and 2.2x faster than Chandra OCR 2. Nine languages hold accuracy. Everything else is zero-shot at lower accuracy, which is Cohere’s phrasing and we appreciate it being said out loud.
When we would reach for it
If you are already paying frontier token rates to turn PDFs into text for a RAG index, run a bake-off. That is the case Parse 5 was built for, and a five point gap on a partial benchmark is cheap next to the invoice difference. Same shape of argument as Gemini 3.5 Transcribe, which also came fifth on quality and won on price.
If your documents are mostly charts, or an auditor needs every number traceable to a coordinate, do not take 79.2 as the answer. Test the two dimensions Cohere left out, on your own pages.
And if you needed open weights, this is not that release. The base architecture is on Hugging Face. The parser is a service.
Nils Reimers, who runs AI Search at Cohere, put the framing in one line that we think is right even if the launch numbers are selective:
Document parsing isn’t solved because the hard part isn’t reading text, it’s preserving structure and meaning.
Sources
Official announcement: Introducing Parse: Enterprise document intelligence at scale and the Parse product page. Independent coverage and the Reimers quote: VentureBeat and MarkTechPost. Benchmark definition and the grounding figures: ParseBench: A Document Parsing Benchmark for AI Agents and the run-llama/ParseBench repository. Availability on Azure: Microsoft Foundry. The cost arithmetic is our own, worked from Cohere’s published per-page price and its 13 million page example, on 31 August 2026.
Frequently asked questions
How much does Cohere Parse 5 cost?
$1.50 per 1,000 pages through the Cohere API, which is $0.0015 a page. Cohere also sells it through Model Vault, its single-tenant deployment, where it claims 23% savings against the API at 50% GPU utilisation and up to 61% at full hourly utilisation. Parse 5 is on Microsoft Foundry and AWS SageMaker too, and Cohere will deploy it in a private cloud or on-premises.
What is the ParseBench score of Parse 5?
Cohere reports 79.2, broken down as 87.0 on tables, 86.6 on content faithfulness and 64.0 on semantic formatting. Those three average to exactly 79.2, so the charts and visual grounding dimensions are not folded into that number. ParseBench itself is a five dimension benchmark from LlamaIndex, published as arXiv 2604.08538, built on roughly 2,000 human-verified pages from insurance, finance and government documents.
Is Parse 5 better than GPT-5.5 or Claude Opus at reading documents?
No, and Cohere's own table says so. On the three dimensions Cohere reports, GPT-5.5 scores 84.4 and Claude Opus 4.8 scores 84.3 against 79.2 for Parse 5. Gemini 3.5 Flash sits at 81.8. The pitch is not that Parse 5 wins, it is that you pay a per-page rate for a 2.3B model instead of frontier token pricing for a job that is mostly mechanical.
Are the Parse 5 weights open?
No. Parse 5 is served through the Cohere API, Model Vault, Microsoft Foundry and AWS SageMaker. The architecture it is built on, CohereLabs/North-Micro-Vision-Instruct, is published on Hugging Face, and there is a demo space, but the parse-v5.0 checkpoint itself is not a download.
What formats and languages does Parse 5 handle?
You send a PDF, PPT or JPEG page as a base64-encoded data URI and get back Markdown: text in reading order, tables rendered as HTML, lists, form key-value pairs, image descriptions and bounding box coordinates. Nine languages hold their accuracy (Arabic, English, French, German, Italian, Japanese, Korean, Portuguese, Spanish) with zero-shot support elsewhere at lower accuracy, in Cohere's own words. The context window is 8,192 tokens.