Cohere Parse 5 at $1.50/1,000 pages: reprice your intake

Cohere shipped Parse 5 on August 28, a 2.3-billion-parameter vision language model that turns PDFs, slides, and images into structured Markdown. It is not the most accurate parser available, and Cohere is not pretending otherwise. It is priced at $1.50 per 1,000 pages, and for anyone running document intake at volume, that is the number that decides the architecture.

What actually happened

Per Cohere's own announcement and the model documentation, parse-v5.0 is a 2.3B-parameter VLM with an 8,192-token context window. It accepts PDF, PPT, and Base64-encoded JPEG, and returns Markdown containing document text, lists, tables rendered as HTML, bounding-box coordinates, and image descriptions. Nine languages are stable — Arabic, English, French, German, Italian, Japanese, Korean, Portuguese, Spanish — with zero-shot attempts at others at lower accuracy.

The benchmark story, via VentureBeat's coverage: Parse 5 scores 79.2 on ParseBench across tables, content faithfulness, and semantic formatting. That trails GPT-5.5 (84.4), Opus 4.8 (84.3), and Gemini 3.5 Flash (81.8). It beats the specialized parsers it is aimed at.

Throughput is 4.5 pages per second. Deployment runs through the Cohere API, Model Vault (single-tenant managed inference, which Cohere claims cuts cost up to 61% at full utilization), Microsoft Foundry, AWS SageMaker, and private environments. Cohere puts the API price at more than 98% below GPT-5.5 for large-scale document workflows.

Why cost per page matters for your business

Here is the thing about document intake: the accuracy you need is set by your review step, not by a leaderboard.

Run 200,000 supplier invoices a year. At $1.50 per 1,000 pages that is $300 of parsing. The same volume through a frontier general-purpose model lands in five figures. If the frontier model gets you from 79 to 84 on a benchmark and a human still confirms totals before anything posts to the ledger, you paid a lot for a number nobody in your business reads.

The inverse is also true. If parsing output goes straight into a payment run with no review, five points of faithfulness is not a rounding error — it is a wire to the wrong account. Pick the tier that matches what happens downstream, not the one that wins the eval.

Two things we would check before switching. Test on your ugly documents — ParseBench is not your fax-scanned packing slip off a 2009 template, so diff 100 of your worst PDFs. And tables as HTML is the real feature: line items are the whole job in commerce and finance intake, and a parser that flattens a table into prose has lost the data no matter how it scored on text.

The private-deployment path is what we weight heaviest. A 2.3B model with a ~4.6GB footprint runs on hardware you can buy, which means your document pipeline is not hostage to one vendor's pricing page. Same argument we made for Mistral OCR 4 and DeepSeek's token-per-image repricing. Document parsing is commoditizing fast, and the teams who kept the parser behind their own interface take each price cut without a rewrite.

Sources: Cohere — Introducing Parse, Cohere Docs — Parse, VentureBeat.