Cohere Parse costs $1.50 per 1,000 pages. Check the facts it extracts.

2026-08-28

Maestro’s take

Before paying for a smarter chatbot, check whether it can read your documents correctly.

A model cannot reliably answer “What is the total?” if the earlier extraction step moved a number into the wrong row. Cohere’s new Parse model makes that less glamorous step worth a fresh look.

TL;DR

Cohere launched Parse on August 27. It turns document pages into Markdown for search and AI applications. Its direct API price is $1.50 per 1,000 pages. Cohere says it is generally available.

That works out to:

These are simple calculations at Cohere’s published API rate—not the total cost of an AI app. Page conversion, storage, search, answer generation, repeat processing, and any human checking are separate. Original launch and pricing.

Why Maestro users care

Suppose you are building a portal that answers questions about supplier invoices.

One approach sends every page to a large model whenever somebody asks a question. Another extracts the pages once, stores the results, then gives the answering model only the relevant material.

Our take: test the second approach when documents are reused. The saving depends on your current workflow. If you already extract and cache documents, there may be much less to gain.

Keep a link from each extracted fact back to its source page. Cheap answers are not much use if nobody can check them.

What it does—and does not—give you

For a direct API integration, plan to send page images. The API reference currently accepts image URLs or embedded image data—not PDF/file URLs. Convert PDF pages first. It offers Markdown or content blocks, not a ready-made invoice record with your chosen fields. You still need validation and a mapping step. API reference.

The model guide says extracted content has no confidence scores. Do not expect a score to tell you which total is safe to trust. Model guide.

Do not read “understands documents” as “understands everything on a page.” Cohere’s launch comparison excludes chart-data extraction and a full layout-scoring dimension. Its reported results are vendor-run, not a Maestro test. Benchmark scope and footnotes.

Try your ugly documents first

Do: choose 20 redacted or synthetic pages: crooked scans, dense tables, credit notes, and multi-page invoices. Write down the correct totals, dates, and row associations before running the parser.

Do not: judge it by whether the Markdown looks neat. A beautifully formatted wrong total is still wrong.

Why: your application needs the right facts, not a pretty transcription.

Compare it with your existing extractor on those same pages. Measure mistakes and correction time alongside the bill. If the facts improve and the full workflow gets cheaper, expand the trial.

Published August 28, 2026. Maestro’s opinion is AI-generated. Product facts come from the linked Cohere sources.