Cohere Parse v5 in Foundry: document parsing cost math

On 27 August 2026 Cohere released Parse v5, a 2.3-billion-parameter vision language model that turns scanned pages into Markdown, and Microsoft listed it in the Foundry model catalog the same day as Cohere-parse-v5 (Preview). The price is $1.50 per 1,000 pages. That is the same list price Azure charges for plain OCR with Document Intelligence Read, and roughly a seventh of what Document Intelligence Layout costs for table-aware extraction. For any Swedish or EU team feeding scanned contracts and invoices into a RAG index on Azure, the arrival of a third serious document parser next to Microsoft's own services and Mistral OCR 4 changes the per-page arithmetic. It also comes with two catches that matter more here than in the US: the model's language list does not include Swedish, and the Foundry deployment is still in preview without a published EU Data Zone option.

What Parse v5 is, and what it is not

Parse is a purpose-built parser, not a general chat model. According to Cohere's documentation the model is 2.3B parameters (about 4.6 GB on disk) with an 8,192-token context, served through a POST /v2/parse endpoint that takes one document at a time as a base64 data URI and returns either Markdown or a list of typed blocks. Tables come back as HTML inside the Markdown so merged cells survive; images get a text description; every block carries bounding-box coordinates so you can highlight the source region in a viewer. Cohere reports a throughput of 4.5 pages per second per GPU, or about 36 pages per second on an 8×H100 node under vLLM.

The limitations are documented just as plainly. Parse returns no confidence scores. It does not extract fields against a schema; you get the page as Markdown and bring your own extraction step. It cannot tell a running header from body text or recover font hierarchy. The API accepts one file per call, capped at 20 MB compressed or 50 megapixels decoded. And the model was trained on nine languages: Arabic, English, French, German, Italian, Japanese, Korean, Portuguese and Spanish. The Foundry catalog entry adds Simplified Chinese to that list. Swedish, Danish, Norwegian and Finnish are absent.

Cohere published its own benchmark, ParseBench, scoring tables, content faithfulness and semantic formatting while explicitly leaving out layout and chart dimensions. On that benchmark Parse 5 scores 79.2, behind GPT-5.5 (84.4), Claude Opus 4.8 (84.3) and Gemini 3.5 Flash (81.8), and ahead of Mistral OCR 4 (74.5), Databricks AI Parse (72.4) and Azure Document Intelligence (69.3). Treat those numbers the way you would treat any vendor benchmark: as a claim to verify on your own documents, not a ranking to act on. The interesting part is not that a 2.3B model loses to frontier models by five points. It is that it does so at a per-page price well over an order of magnitude lower, which is exactly the trade most document pipelines want.

The four options on Azure, priced in Sweden Central

All figures below are USD pay-as-you-go list prices per 1,000 pages, pulled from the Azure Retail Prices API for the Sweden Central region on 31 August 2026, plus the vendors' own published rates where a model is billed directly. Your EA or MCA discount applies on top; none of these numbers include the compute you run around the parser.

Two things stand out. Parse v5 is priced at the OCR tier while delivering layout-tier output. And Microsoft's own layout-aware option is not Document Intelligence Layout at $10 any more; it is Content Understanding's Standard meter at $5, which is the same engine surfaced through the newer service. If you are still calling prebuilt-layout on the classic endpoint for RAG ingestion, you are paying double for the same extraction before Parse even enters the picture.

Cost math at three volumes

Assume scanned, layout-heavy pages (invoices, agreements, claim forms) where you want tables preserved and Markdown for an index. Monthly cost, list prices, Sweden Central or the vendor's Global rate as noted.

At 100,000 pages a month the gap between Parse and Content Understanding is $350. Nobody should re-platform a working pipeline for $350. At a million pages the gap is $3,500 a month, which pays for the evaluation project but not much else. At ten million pages the difference is $35,000 a month, or $420,000 a year, and the question becomes why you would not at least run a bake-off. The volume threshold where Parse v5 deserves a proper evaluation sits somewhere around one to two million pages a month for most organisations; below that, the switching cost and the preview risk outweigh the saving.

Above that threshold, Cohere would rather sell you dedicated capacity. Its pricing page lists Model Vault instances for Parse 5 at $2,500 a month (Medium) and $4,300 a month (XL), with the company claiming a 23% saving over API pricing at 50% GPU utilisation and 61% at full utilisation. Those numbers are Cohere's, and they assume you keep the box busy around the clock. Note that Model Vault is a Cohere-hosted product, not a Foundry deployment, so it sits outside your Azure tenant and outside the residency argument below.

The frontier-model alternative deserves one sentence. Sending every page to GPT-5.5 or Claude Opus as an image gets you the top ParseBench scores, and Cohere's own modelling puts Parse at a 98% cost reduction against that approach at scale. The sensible pattern is not either/or: parse everything with a cheap specialist, then escalate the small fraction of pages that fail validation to a frontier model.

Where the cheap option loses

Price per page is the wrong metric if the output needs a human review queue. This is where the Microsoft services earn their premium.

• Confidence scores and grounding. Document Intelligence returns confidence per word and per field. Content Understanding adds opt-in source grounding for every extracted field, so you can route anything below a threshold to a reviewer and let the rest flow. Parse gives you neither. If your process depends on auto-approving 80% of invoices and reviewing 20%, you need a signal Parse does not emit, and you will end up building a second pass to generate one.
• Field extraction. Parse stops at Markdown. Content Understanding's prebuilt invoice, receipt, contract and procurement analyzers go straight to typed fields, and custom analyzers let you label a handful of samples for your own forms. Add the $1.00 contextualization meter plus model tokens, and you are still comparing a finished extraction against a raw page dump.
• Handwriting and selection marks. Document Intelligence's Read and Layout models handle handwritten text in twelve languages and detect checkboxes and radio buttons. Parse's documentation makes no handwriting claim.
• Request shape. Parse takes a single image or file up to 20 MB per call and no file URLs. A 400-page scanned agreement means splitting and reassembling on your side. Document Intelligence and Content Understanding accept multi-page PDFs and blob URLs directly.
• Support status. Cohere-parse-v5 is marked Preview in the Foundry catalog, with no SLA. Document Intelligence and Content Understanding are GA services with the standard Azure SLA. For Swedish public-sector procurement, that alone can be disqualifying.

Calling it: Parse versus Layout side by side

The integration surface is small either way. Cohere's SDK call, which Foundry exposes with the same model id, looks like this:

import base64, cohere

co = cohere.ClientV2()  # or the Foundry endpoint + key

with open("faktura-2026-08.pdf", "rb") as f:
    data_uri = "data:application/pdf;base64," + base64.b64encode(f.read()).decode()

response = co.parse(
    model="parse-v5.0",
    document={"type": "image_url", "image_url": data_uri},
    output_format="blocks",
)

for page in response.pages:
    for block in page.blocks:
        if block.type == "table":
            print(block.table.bounding_box)
            print(block.table.html)

The equivalent Document Intelligence Layout call with Markdown output (v4.0 API) is:

from azure.ai.documentintelligence import DocumentIntelligenceClient
from azure.ai.documentintelligence.models import AnalyzeDocumentRequest, ContentFormat
from azure.identity import DefaultAzureCredential

client = DocumentIntelligenceClient(endpoint, DefaultAzureCredential())

poller = client.begin_analyze_document(
    "prebuilt-layout",
    AnalyzeDocumentRequest(url_source=blob_sas_url),
    output_content_format=ContentFormat.MARKDOWN,
)
result = poller.result()
print(result.content)          # Markdown, tables as HTML
for page in result.pages:
    for word in page.words:
        if word.confidence < 0.9:
            flag_for_review(word)

Both return Markdown with HTML tables, which makes a fair comparison straightforward: run the same 200 documents through each, diff the tables, and count the pages where a human would have to intervene. That count, multiplied by your reviewer's hourly cost, is the number to put next to the per-page saving.

The Swedish and EU angle

Language coverage is the first gate. Parse's trained languages do not include Swedish or any other Nordic language. A Latin-script OCR model will usually read å, ä and ö well enough, but "usually" is not a vendor commitment, and compound Swedish invoice vocabulary, personnummer formats and Bankgiro references are exactly where a model trained elsewhere drifts. Document Intelligence lists Swedish explicitly for printed text in both Read and Layout, Content Understanding advertises text extraction across hundreds of languages, and Mistral claims 170 languages for OCR 4. Before any cost conversation, run 50 real Swedish documents through Parse and measure character error rate on the fields you care about. If it holds, the price is compelling. If it does not, the price is irrelevant.

Residency is the second gate. Microsoft's announcement says Parse inference "stays within your Azure tenant boundary" and inherits your network, identity and data governance controls. That is true of every model sold directly by Azure and is a real advantage over calling Cohere's API or Model Vault, where the data leaves Azure entirely. What Microsoft has not yet published, as of the 27 August update to the Foundry model pages, is which deployment types and regions Parse supports. Other Cohere models in Sweden Central are billed on both Global and EU Data Zone meters, so a Data Zone option is plausible, but until it appears in the region tables you should assume Global Standard, which by Microsoft's definition may process requests in any Azure region worldwide. For personal data under GDPR that is a transfer-impact question, not a checkbox. Mistral OCR 4 already has an EU Data Zone meter in Sweden Central at $4.40 per 1,000 pages, and Document Intelligence and Content Understanding are ordinary regional resources. If EU-only processing is a hard requirement today, Parse is the option to watch rather than the option to deploy.

Traceability helps with the AI Act. Document parsing itself is not a high-risk use under Annex III, but the systems it feeds often are, credit decisions and insurance claims handling among them. When the December 2027 high-risk obligations arrive, you will need to show where an extracted value came from. Bounding boxes from Parse or Mistral, and source grounding from Content Understanding, are the cheapest way to build that audit trail now rather than retrofit it later. A parser that returns text without coordinates is a liability in that context regardless of price.

A decision guide

Cohere has priced Parse v5 to make the per-page line item disappear from the conversation, and at volume it does. What it has not yet done is publish the language coverage and residency options a Swedish enterprise needs before that line item is the only thing left to discuss. Test it this quarter; deploy it when the catalog entry loses the Preview tag.

Sources

• Cohere: Introducing Parse, enterprise document intelligence at scale (27 August 2026)
• Cohere docs: Parse model overview, languages and limitations
• Cohere API reference: POST /v2/parse, request limits
• Cohere pricing: Model Vault instance rates for Parse 5
• Microsoft Tech Community: Cohere Parse v5 is now available in Microsoft Foundry
• Microsoft Learn: Foundry Models sold by Azure (Cohere-parse-v5 and mistral-ocr-4-0 entries)
• Azure Retail Prices API (Document Intelligence, Content Understanding and Mistral meters, Sweden Central)
• Microsoft Learn: Pricing for Azure Content Understanding
• Microsoft Learn: Document Intelligence Layout model and Markdown output
• Microsoft Learn: Document Intelligence language support for Read and Layout
• Microsoft Tech Community: Mistral Document AI (OCR 4) arrives in Microsoft Foundry
• Mistral AI: OCR 4 announcement
• VentureBeat: Cohere Parse 5 loses the benchmark on points, wins on cost per page (ParseBench table)