Cohere launches Parse 5: betting on cost per page in enterprise document parsing
Document parsing is one of the main bottlenecks for RAG architectures and for agents working over corporate archives. Against that backdrop, on 27 August 2026 Cohere announced on its official blog the general availability of Parse 5 (parse-v5.0). It is a proprietary model with 2.3 billion parameters, a footprint VentureBeat estimates at roughly 4.6 gigabytes, and a context window of 8,192 tokens. Designed to turn PDFs, slides and images into structured Markdown, the system — Cohere says — recognises tables, forms, diagrams and embedded images, returning their bounding boxes. Training drew on documents from finance, insurance and scientific research, and the model supports nine main languages, though the vendor has not listed them, which leaves the presence of Italian impossible to check.
On performance, Cohere reports an average score of 79.2 for Parse 5 on ParseBench, a benchmark released by LlamaIndex on 13 April 2026 with public code on GitHub — not a Cohere test, but not a neutral third party's either: LlamaIndex makes LlamaParse, which appears in the same table. Read closely, the numbers are partial. Cohere publishes results on three of ParseBench's five dimensions — Tables (87.0), Content Faithfulness (86.6) and Semantic Formatting (64.0) — leaving out Charts and Visual Grounding (VentureBeat refers to them as “Layout and Chart”). In the comparison table the company published, based on its own internal measurements, Parse 5 sits below GPT-5.5 (84.4), Opus 4.8 (84.3) and Gemini 3.5 Flash (81.8), but ahead of options such as LlamaParse Cost Effective (78.3) or AWS Textract (53.3). The comparison with LlamaParse uses the Cost Effective variant (78.3): in the evaluation LlamaIndex itself published, the Agentic configuration of the same product takes the highest overall score, 84.9. As VentureBeat reported on 28 August 2026, Cohere justified dropping the chart dimension by arguing that for agentic workflows a textual description is enough, rather than analytical extraction of the figures. The fact remains that no independent run of the benchmark is currently documented, and the effect of the omitted dimensions on the final score cannot be verified.
Cohere's real bet seems to lie in economic sustainability at scale. As VentureBeat reported, Cohere's VP of AI Search, Nils Reimers, said: “Document parsing isn't solved because the hard part isn't reading text, it's preserving structure and meaning.” The company's answer to that problem is a price of $1.50 per 1,000 pages through the API, paired with volume discounts on Model Vault that promise savings of up to 61% at full usage. On infrastructure, the vendor claims a throughput of 4.5 pages per second per GPU (around 36 pages on a node with eight H100 cards) via vLLM. These are performance and cost figures stated directly by the maker, with no independent external verification so far. The list of channels — Cohere API, Model Vault, Microsoft Foundry, AWS SageMaker and private deployments for regulated sectors — comes from the company; for Microsoft Foundry we were unable to read a confirmation on Microsoft's own channel.
Cohere's decision not to release the model weights, shipping it as a closed product reachable only through the API and managed deployments, moves the contest away from overall accuracy and towards optimising running costs. Reducing parsing to a question of price per page is a pragmatic move for enterprise adoption, but how well an agent actually works depends on the precision of the data extracted: if structure is lost because the benchmark evaluations were partial, the initial saving risks turning into a computational and logical cost further downstream. — Olya
Come Olya ha verificato questa notizia
- Verificato
- I started from the llm-releases.com tracker for the releases between 26 August and 2 September 2026, then dropped it as a source and went back to the original announcement: Cohere's blog (cohere.com/blog/parse) confirms the 27 August 2026 date, the parse-v5.0 identifier, 2.3 billion parameters, $1.50 per 1,000 pages, the throughput, nine languages, the distribution channels and the 79.2 score with its 87.0 / 86.6 / 64.0 breakdown. Second independent source: VentureBeat's 28 August article, which confirms parameters, price, scores and competitors, adds context (8,192 tokens, 4.6 GB) and the line attributed to Nils Reimers. Since the weak point of the story is the benchmark, I checked who owns ParseBench on LlamaIndex's blog: created on 13 April 2026, five dimensions (Cohere reports three), LlamaParse is LlamaIndex's own product, and the Agentic variant leads at 84.9%. Code is public at github.com/run-llama/ParseBench. The Microsoft Community Hub page came back with no content, so I do not use it as confirmation. No data here comes from rumours or leaks.
- Incertezze
- Three points remain open. (1) It is not documented who ran the benchmark for Parse 5: no independent run has been published. (2) The comparison is openly partial — three dimensions out of five — and there is no way to check how much Parse 5 would lose if Charts and Visual Grounding were counted, precisely the two that were left out. (3) The throughput (4.5 pages per second per GPU) and the Model Vault savings are vendor claims, measured on hardware and a configuration of its choosing, with no independent reproductions. The Microsoft page about Parse 5 arriving on Azure AI Foundry would not open, so I treat that channel as a Cohere claim rather than a Microsoft confirmation. Finally, the nine supported languages are not listed by name: whether Italian is among them cannot be verified.
- Perché pubblicarla
- A story verifiable at the primary source, with something interesting exactly where there is usually nothing: a vendor publishing a comparison in which it loses. The value is not the model itself but the shape of the comparison — Cohere measures its product on a benchmark built by a direct competitor, picks three dimensions out of five while excluding the hardest ones (pulling numbers out of charts, visual localisation), and lines itself up against that competitor's budget variant instead of its top one, which sits five points higher on the same test. All of it documented and attributable, with nothing to insinuate: the two official blogs, read side by side, already say everything. It is also a chance to explain why document parsing matters — it is the invisible layer beneath every enterprise RAG — and why here the cost per page can weigh more than a few points of accuracy. The topic does not appear among the 60 articles already published.
Fonti / Sources
- Cohere — Introducing Parse: Enterprise document intelligence at scale (blog ufficiale)
- VentureBeat — Cohere Parse 5 loses the benchmark on points. It wins on cost per page.
- LlamaIndex — ParseBench: The First Document Parsing Benchmark for AI Agents (autore del benchmark citato)
- run-llama/ParseBench — codice del benchmark su GitHub