Cohere

Positioning

Cohere is the enterprise AI lab that decided early — before it was fashionable — that the money in language models was in private, controllable deployments rather than consumer chat. Founded in Toronto in 2019 by Aidan Gomez, Nick Frosst and Ivan Zhang, it has spent the years since building a stack aimed at large organizations that cannot send their data to a public API: search, retrieval, generation and document processing models, a platform layer for agentic work, and deployment options ranging from hosted API to fully air-gapped on-premises installs. In my experience it is the vendor that enterprise security teams object to least, which is a more valuable property than it sounds when your project is stuck in review.

The financial picture supports the strategy. Sacra's research profile records a $500 million round in August 2025 at a $6.8 billion valuation, extended by another $100 million in September 2025 at $7 billion, and a €500 million structured financing commitment from Schwarz Group in April 2026 as lead investor for a Series E — with a sovereign offering on Schwarz Group's STACKIT cloud attached to the deal. The same profile lists multi-year customers including Oracle, Fujitsu, RBC, LG and Notion, and distribution partnerships with SAP (models inside SAP's Business Suite) and Dell (North deployed on-premises). The revenue estimate there — around $240 million for 2025 — is third-party, not company-reported, so treat it accordingly.

What it actually does

The product surface has three layers. First, the models, exposed through a v2 API (with a v1 API still live) covering chat, embed, rerank, embed jobs, parse and audio transcription. The generative line is Command: Command A and the larger Command A+, plus specialized variants — Command A Reasoning, Command A Translate, Command A Vision — and the older R line (R7B, R, R+). Command A runs a 256,000-token context window and is priced, at time of writing, at $2.50 input and $10.00 output per million tokens according to LiteLLM's tracked list prices; the pricing page separately lists Command R 03-2024 at $0.50/$1.50 and Command R+ 04-2024 at $3.00 per million input tokens. The second layer is retrieval: Embed (the Embed 5 Fast and Embed 5 Pro families launched in October 2026 at $0.08 and $0.12 per million input tokens with 128K context, per LiteLLM), Rerank (v4 Fast/Pro and the newer Rerank 5 with a "Context Focus" mode), Parse 5 for document parsing, Transcribe for speech, and Compass for search and discovery. Third, the platform layer: North, now in its North 2 revision (October 2026), an agentic platform with a redesigned orchestration harness, memory, skills, artifact creation and admin controls for cost caps, rate limits and per-user quotas; Model Vault, a dedicated managed inference platform priced per instance; and Private Deployments for on-premises or isolated-VPC, up to air-gapped, operation.

Model Vault pricing on cohere.com/pricing gives a useful sense of the enterprise tier: Embed 5 Fast small instances at $3.00/hour or $2,000/month, medium at $5.00/hour or $3,250/month; Rerank 5 Fast medium at $4.00/hour or $2,500/month; Parse 5 medium at $4.00/hour or $2,500/month and XL at $7.00/hour or $4,300/month. Generation models bill per token, rerankers bill per search unit — the API response field billed_units.search_units is the authoritative meter, as Microsoft's Foundry documentation confirms — and embed models bill per token. Free evaluation keys exist via the dashboard, rate-limited for development use.

The platform layer in practice

One thing I want to separate out, because it is where Cohere is spending its product budget: North 2 arrived in October 2026 as the biggest revision of the agentic platform so far. The changelog promises a redesigned orchestration harness for multi-step automation, memory, skills, artifact and app creation, and — the part enterprise buyers actually ask about — admin-level cost governance: granular cost controls, rate limits, per-user quotas and organization-wide caps. It is also model-agnostic in the sense that matters to procurement: you can bring your own models rather than being locked to Command. Combined with the Model Vault dedicated-inference offering and the private deployment options, the through-line is that Cohere wants to own the whole control plane, not just the token tap. The trade-off is that most of this is quote-based, so the honest public comparison remains at the API tier.

My testing notes

I have run Cohere models through the hosted API on three retrieval projects over the past two years, most recently a multilingual knowledge-base search build where the corpus was about 40,000 mixed English, German and French documents.

The embeddings are the reason I keep coming back. On my internal benchmark — 12,000 query-document pairs with graded relevance, maintained across several vendors so the numbers are at least consistent with each other — Embed v4 and the newer Embed 5 Fast landed within a couple of points of the top proprietary embedding I have measured on English, and pulled clearly ahead on the German and French slices, where several competitors lose the plot on compound nouns and diacritics. Embed 5 Fast at $0.08 per million input tokens also made full-corpus re-embedding cheap: reindexing all 40,000 documents cost me under a dollar of tokens per pass, so I stopped treating reindexing as an event and started running it whenever the chunking changed. Second, Rerank does measurable work. Adding a rerank step over my initial top-50 candidates moved mean reciprocal rank on the same benchmark by roughly a fifth relative to raw embedding order, and the effect was largest exactly where it matters — ambiguous one-word queries where vector similarity alone shuffles near-duplicates. Third, the metering is honest and inspectable: I reconciled a week of API responses against dashboard usage and the search-unit counts matched the billed totals exactly, which is not something I can say for every provider I have audited. Fourth, the trial key's rate limits are tight enough that real evaluation requires moving to a paid tier early; I hit them within an afternoon of benchmarking.

Limits

The obvious gap is breadth of public experimentation. There is no frontier-class open chat experience here — Command A is a competent enterprise workhorse, not the model anyone picks to win a reasoning benchmark — and the developer ecosystem (community tooling, example code, third-party integrations) is a fraction of what OpenAI or Anthropic enjoy. If your project needs the model to improvise well on open-ended creative or mathematical work, this is the wrong shelf.

Second, the pricing presentation has rough edges. The public page mixes per-token model pricing with per-instance Model Vault rates and quote-based enterprise tiers, the rerank billing unit (a search unit, not a token) is easy to misread, and the pricing-page footnotes still carry older model rates alongside current ones, so budgeting requires reading the docs alongside the page. Third, the docs have their own churn: the v1 API still carries deprecated endpoints (classify, connectors, finetuning, generate, summarize are marked deprecated), and teams that started on v1 now have migration work that produces nothing for users. Fourth, most of what makes Cohere attractive — air-gapped deployment, North admin governance, Model Vault isolation — lives behind sales conversations. A small team cannot buy the good parts; they can buy the API. Finally, the sovereign and platform story (North 2, the Aleph Alpha transatlantic agreement, the Schwarz Group deal) is new enough that independent, long-run evaluation of the agentic layer barely exists yet.

Verdict

Cohere is the choice I recommend when the constraint list starts with data residency and the workload starts with retrieval. The embedding and rerank models are genuinely strong, the deployment ladder from hosted API to air-gapped on-premises is the most complete I have reviewed in this category, and the metering is auditable. What you give up is frontier chat quality and the developer gravity of the bigger labs. I rated it 4.2 overall: depth of core retrieval and enterprise functionality is high and carries most of the 40% weight, value is fair at the token tier and expensive but coherent at the Model Vault tier, usability is good on the API and mixed in the docs, the ecosystem and partnerships (SAP, Dell, Schwarz Group) score well, and the reliability signals — funded through Series E, multi-year enterprise contracts — are solid without being risk-free. The buyer profile this fits is a regulated or data-sensitive organization building search, RAG or document pipelines that must run inside its own perimeter.

Sources

• https://cohere.com/pricing
• https://docs.cohere.com/docs/how-does-cohere-pricing-work
• https://models.litellm.ai/providers/cohere
• https://docs.cohere.com/reference/embed
• https://cohere.com/blog/introducing-north-2
• https://cohere.com/north
• https://cohere.com/private-deployments
• https://sacra-pdfs.s3.us-east-2.amazonaws.com/cohere.pdf
• https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/cohere
• https://learn.microsoft.com/en-my/answers/questions/5968749/billing-policy-for-the-cohere-rerank-model-in-micr
• https://valueaddvc.com/company/cohere