Cohere

Editor’s note: This Cohere review is based on my hands-on test of Cohere’s API with Embed 5, Rerank 5 and Command A+, on a trial key and a production key, in November 2026. I weighed pricing, limits, retrieval, generation and deployment against OpenAI, Voyage AI, Google and Mistral.

Quick verdict: I recommend Cohere for RAG and search teams that want embeddings, reranking and generation from one vendor, plus a private or air-gapped deployment path. The trial key bars commercial use, and for plain text, OpenAI’s text-embedding-3-small costs a quarter of Embed 5 Fast.

Embed 5 fuses text and page images into one vector with a 128K-token context, earning my recommendation in this Cohere review for visually rich document search. In this review, I’ll take a closer look at Cohere’s pricing, features, and deployment options, so you can see exactly whether it fits your search stack or whether OpenAI or Voyage AI is the better buy.

Key Takeaways

• Index with Pro, query with Fast: Embed 5 Pro ($0.12 per 1M tokens) and Fast ($0.08) share one embedding space.
• Rerank 5 Fast costs $0.02 per 1M tokens: Pro costs $0.05; both return Context Focus passage spans.
• Command A+ weights are Apache 2.0: they fit on two H100s; older Command A, R and Aya weights are noncommercial.
• Trial keys stop at 1,000 calls a month: Command A+, Reasoning, Vision and Translate production limits go through sales.
• Bedrock lacks Embed 5 and Rerank 5: the HIPAA BAA also excludes the SaaS API.

Pros and Cons

Cohere Pricing: How Much Does It Cost?

Every Cohere account starts on a free trial key, and production use is pay-as-you-go per 1M tokens, billed at the end of each calendar month or when the balance reaches $250.

• Trial key (free): for prototyping only, with 1,000 API calls a month, 20 chat requests a minute per model, 10 rerank requests a minute, 2,000 embed inputs a minute and 5 image-embed inputs a minute. Production and commercial use are not permitted.
• Production key (pay-as-you-go): for live apps, once an organization Owner completes Go to Production. Limits rise to 500 chat requests a minute on Command A, R, R+ and R7B, 1,000 rerank requests a minute and 400 image-embed inputs a minute.
• Model Vault (per GPU-hour): for dedicated, Cohere-managed inference with zero data retention by default and no rate limits, from $3.00 an hour for an L4 to $12.50 for a GB200 (H100 $6.25). Promotional rates of $1.00 (L4) and $4.25 (H100) run through March 31, 2027.
• North, Compass and private deployment (custom): for enterprise deployments, quote-only.

Model	Price per 1M tokens	Context	Notes
Embed 5 Pro	$0.12	128K	Text, images and fused inputs
Embed 5 Fast	$0.08	128K	Shares Pro’s embedding space
Rerank 5 Pro	$0.05	32,768	Context Focus spans
Rerank 5 Fast	$0.02	32,768	Calibrated scores and Context Focus
Command A+	$0 on an API key	128K input / 64K output	Production limits through sales
Command R	$0.15 input / $0.60 output	128K	4K maximum output
Command R7B	$0.0375 input / $0.15 output	128K	4K maximum output
Parse 5	$1.50	8K	Markdown and HTML document parsing

Command A, Command A Reasoning, Command A Vision and Command A Translate have no public per-token price. On Command A+, Reasoning, Vision and Translate, production keys stay capped at 1,000 calls a month until sales sets their limits. Transcribe runs from $3.75 an hour per instance on Model Vault.

Is Cohere Good Value for Money?

• Billion-token indexing: $120 on Embed 5 Pro and $80 on Embed 5 Fast, against $130 on OpenAI’s text-embedding-3-large and $20 on text-embedding-3-small. Embed 5 undercuts the large model and adds image and fused-page inputs, which OpenAI’s text-only models do not take.
• Images inside the token price: Embed 5 has no separate per-image charge, while Voyage’s voyage-multimodal-3.5 adds $0.60 per billion pixels to its $0.12 per 1M text tokens.
• Text-only rivals: Voyage AI’s voyage-4 costs $0.06 per 1M tokens with the first 200M tokens free; Mistral’s mistral-embed costs $0.10 with an 8K context.
• Reranking at parity: Rerank 5 Pro and Fast cost exactly what Voyage’s rerank-3 ($0.05) and rerank-3-lite ($0.02) cost per 1M tokens.
• Split-model savings: Pro’s $0.12 rate applies to the one-off indexing job, and every live query runs on Fast at $0.08.

Cohere is good value for multimodal or multilingual retrieval with a reranker from the same vendor. It is poor value for an English-only text app, where text-embedding-3-small costs $0.02 per 1M tokens.

Reviewer’s Notes: I recommend building on the trial key and moving to a production key before any customer traffic, because the trial terms bar commercial use. Index with Embed 5 Pro, query with Embed 5 Fast, start reranking on Rerank 5 Fast, and talk to sales early if Command A+ or Command A Reasoning is your generation model at production volume.

My Experience With Cohere

I built a two-stage retrieval pipeline from separate Embed, Rerank and Chat calls: Embed 5 Pro for the corpus, Embed 5 Fast for queries, Rerank 5 Fast on the candidates and Command A+ for the grounded answer.

Getting My Trial and Production Keys

I registered at dashboard.cohere.com, which opened a personal account with a free, rate-limited trial key on the API keys page. In the Playground at dashboard.cohere.com/playground, Chat with grounding showed the generated search query and its citations.

As the organization’s Owner, I completed Go to Production:

1. I acknowledged the SaaS agreement and terms.
2. I read the model limitations and the data statement.
3. I answered the sensitive-use-case question.

Declaring a sensitive use case can keep the key rate-limited until a safety review of up to 72 business hours. Under Settings > Data Controls, I opted out of training use of my data; SaaS logs are deleted after 30 days.

Embedding Documents With Embed 5

With the Python SDK’s ClientV2, I called embed with model embed-v5.0-pro and input_type search_document for the corpus, then embed-v5.0-fast with search_query for questions, since Fast shares Pro’s embedding space. I set embedding_types, which the v2 API requires, and output_dimension to 1024 instead of the 2,048 default.

input_type is not optional in practice: documents and queries are embedded differently, so a corpus indexed as search_document pairs with questions sent as search_query.

For PDF pages, I fused each page’s text block and screenshot into one vector, with the image sent as a base64 data URL; one call takes up to 96 inputs. OpenAI’s text-embedding-3 models are text only, so the same pages would need text extraction first.

Image inputs are capped at 5 a minute on the trial key and 400 on production, while text embeds get 2,000 inputs a minute on both. Binary output packed each 1,024-dimension vector into 128 bytes.

Reranking Candidates With Rerank 5

I sent the top vector-search candidates to rerank with model rerank-v5.0-fast, the query, the documents and top_n. Records went in as YAML strings with key order preserved (sort_keys=False), because text past max_tokens_per_doc, 4,096 tokens by default, is truncated; I put the fields that matter first.

With relevant_spans switched on (Context Focus, with mode, threshold and max_spans), each result came back with scored character spans of its most relevant passages, so only those spans went on to Command A+. Calibrated 0-1 scores let me set one relevance threshold across queries. I kept Fast on the live path at $0.02 per 1M tokens; Pro costs $0.05, the same as Voyage’s rerank-3.

Grounded Answers With Command A+

I passed the reranked passages as documents to chat with command-a-plus-05-2026, and the response carried citations pointing back to source spans; safety_mode is forced to CONTEXTUAL whenever documents or tools are passed. With tools, the model returned tool calls that my code executed and sent back, strict_tools enforced the tool schemas, and response_format forced JSON-schema output.

I also pointed existing OpenAI SDK code at the Compatibility API by changing the base URL to https://api.cohere.ai/compatibility/v1, which works with the Python, TypeScript/JavaScript and .NET SDKs, with Java and Go in beta. I dropped n, logit_bias and parallel_tool_calls, which it does not support, along with dimensions on embeddings; reasoning effort accepts only none or high.

Reviewer’s Notes: Set input_type and embedding_types explicitly on every embed call, keep Pro for indexing and Fast for queries, and plan around the trial key’s 1,000-call monthly cap before load-testing anything.

Embed 5: Multimodal Embeddings With a 128K Context

Embed 5 shipped on September 30, 2026 as embed-v5.0-pro (about 2.4B parameters) and embed-v5.0-fast (about 1B), sharing one embedding space across 100+ languages. Cohere’s split puts Pro on offline indexing and Fast on interactive search and agent loops.

• Inputs: text, images and fused text-plus-image inputs such as PDF page screenshots, so visual pages skip text extraction.
• Context: 128K tokens per input, against 512 on Embed v3.
• Matryoshka dimensions: 256, 512, 768, 1,024, 1,536 or 2,048, with 2,048 as the default.
• Output types: float, int8 and binary; a 1,024-dimension binary vector is 128 bytes.
• Batch size: up to 96 inputs per call, with images sent as base64 data URLs.
• input_type values: search_document, search_query, classification, clustering and image.

Embed v4 (128K context, 256 to 1,536 dimensions) and Embed v3 (512-token context) stay live, and Embed v2 was retired on April 4, 2026. The batch Embed Jobs API still only runs Embed v3, so an Embed 5 backfill goes through the regular endpoint at 2,000 inputs a minute, and image inputs at 400 a minute on a production key.

Embed 5 runs on the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker and OpenRouter, but not on Amazon Bedrock, where Cohere’s embeddings stop at Embed v3, as they do on OCI.

OpenAI’s text-embedding-3 models are text only, Voyage’s voyage-multimodal-3.5 bills pixels on top of tokens, and Google’s gemini-embedding-2 goes further than Embed 5 by adding audio and video, at $0.20 per 1M text tokens against Embed 5 Pro’s $0.12.

I recommend Embed 5 for visually rich, multilingual or long-document corpora, indexed with Pro and queried with Fast. Skip it if the whole job is a text-only English index that has to stay on Bedrock.

Rerank 5: Calibrated Scores and Context Focus

Rerank 5 re-scores up to 10,000 documents per request with a 32,768-token context shared by the query and each document. It shipped on October 21, 2026 as rerank-v5.0-pro (about 2B parameters) and rerank-v5.0-fast (about 500M).

• Calibrated 0-1 scores: one threshold decides which documents reach the LLM, whatever the query.
• Context Focus: the API-only relevant_spans option returns scored character spans of the most relevant passages in each document, in the same call.
• Truncation: max_tokens_per_doc defaults to 4,096 tokens, and longer documents are cut.
• Structured records: JSON records and tables go in as YAML strings with key order preserved, so the fields placed first survive truncation.
• Languages: one model covers 100+ languages.
• Text only: no image input, so pages Embed 5 retrieved from screenshots need a text version for reranking.
• Fixed relevance: instructions cannot change how it judges relevance.
• Older models: Rerank 4 and Rerank 3.5 stay live, and Rerank 3.5 is the only Cohere reranker on Amazon Bedrock and OCI.

I recommend adding Rerank 5 behind an existing retriever before replacing the index: it reads candidate text at query time, so it needs no re-embedding. Cohere’s Embed and Rerank integrations cover:

• Elasticsearch, OpenSearch and Vespa.
• MongoDB and Redis.
• Pinecone, Weaviate and Qdrant.
• Milvus and Chroma.
• LlamaIndex, LangChain and Haystack.

Voyage’s rerank-3 and rerank-3-lite match Rerank 5 Pro and Fast at $0.05 and $0.02 per 1M tokens, so the choice follows whichever vendor’s embeddings and deployment options you already use.

Command A+, North and the Rest of the Lineup

Command A+ (command-a-plus-05-2026, released May 20, 2026) is Cohere’s first mixture-of-experts model, with 218B total and 25B active parameters, text and image input, 128K input and 64K output, and 48 languages. Its Apache 2.0 weights fit on one B200 or two H100 GPUs (W4A4), and Cohere calls it the last model in the Command A family.

• Command A: 256K context, 8K output, text only.
• Command A Reasoning: 256K context, 32K output, 23 languages, reasoning on or off.
• Command A Vision: text and image input, 128K context.
• Command A Translate: 23 languages, 8K input and 8K output.
• Command R and R7B: 128K context and 4K output, at $0.15/$0.60 and $0.0375/$0.15 per 1M input/output tokens.
• Parse 5: $1.50 per 1M tokens to turn documents into Markdown or HTML with tables, bounding boxes and image descriptions.
• Transcribe: speech-to-text with Apache 2.0 weights.

Command A+, North Mini Code and Transcribe weights are Apache 2.0, but Command A, A Reasoning, A Vision, A Translate, R, R+, R7B and Aya weights are CC-BY-NC 4.0, so commercial self-hosting of those needs a Cohere agreement.

North 2, the latest release of Cohere’s agent workspace, covers:

• Agents, automations and skills.
• Libraries, applications and memory.
• Guardrails and admin.
• Connectors to Google Drive, SharePoint, Slack, Salesforce and MCP servers.

North runs in a VPC, on-premises (including air-gapped) or on Model Vault, with quote-only pricing, and Compass (search connectors, document parsing and a managed index) is quote-only too. Self-serve fine-tuning, the Classify endpoint, /v1/generate and v1 connectors were retired on September 15, 2025. For commercial self-hosting on open weights, start with Command A+ and skip the CC-BY-NC models.

Deployment, Security and Data Handling

Cohere models run as SaaS, on Amazon Bedrock, Amazon SageMaker, Azure AI Foundry, Oracle OCI Generative AI and OpenRouter, on Model Vault, or privately in a VPC on AWS, Azure, GCP or OCI, on-premises and air-gapped, typically on Kubernetes. Availability differs by channel: Embed 5, Rerank 5 and Command A+ are all missing from Bedrock, and Command A+ is on Azure Foundry and OpenRouter but not OCI.

Deployment	Data handling	Important constraint
Cohere SaaS	30-day log deletion; training opt-out	Zero retention requires enterprise approval
Bedrock, SageMaker, Azure, OCI	Cohere receives no prompts or outputs	Available model versions differ
Model Vault	Zero retention by default; no rate limits	Dedicated GPU-hour billing
Private deployment	Cohere receives no prompts or outputs	Sales-led; infrastructure responsibilities remain

• Model Vault: Standard vaults and Encrypted confidential-computing vaults with zero data retention by default and no rate limits, in US, EU and Canada regions; Encrypted vaults start in the US, with EU and Canada coming. It also hosts rival open models such as DeepSeek, Qwen, Gemma, GPT-OSS and Mistral.
• GPU-hour billing: Model Vault bills per GPU-hour instead of per token and has no rate limits, providing a route past the 500-requests-a-minute chat limit on production keys.
• SaaS data: logs are deleted after 30 days, training use is an opt-out under Settings > Data Controls, and zero data retention is only for approved enterprise customers.
• Compliance: SOC 2 Type II, ISO 27001, ISO 42001 and UK Cyber Essentials, plus GDPR and CCPA.

Cohere signs a HIPAA BAA only for custom model development engagements, so PHI should not go to the SaaS API. The official Python, TypeScript, Java and Go SDKs make the same calls against the Cohere API, Bedrock, SageMaker, Azure and private deployments, and OpenAI SDK code reaches Cohere through the Compatibility API.

Model Vault or a private deployment is my pick for banks, governments and healthcare groups that need models inside their own perimeter.

How Does Cohere Compare to Competitors?

Cohere sells a multimodal embedder, a reranker and open-weight generation models that also run in a VPC, on-premises or air-gapped, and each rival below wins a narrower niche.

Provider/model	Price per 1M tokens	Input coverage	My shortlist reason
Cohere Embed 5 Fast / Pro	$0.08 / $0.12	Text, images, fused inputs	Shared Pro/Fast space
OpenAI text-embedding-3-small / large	$0.02 / $0.13	Text	Low-cost small model
Voyage AI voyage-4	$0.06	Text	First 200M tokens free
Google gemini-embedding-2	$0.20 text / $0.45 image	Text, images, audio, video	Broader media inputs
Mistral mistral-embed	$0.10	Text	Mid-priced text, 8K context

• OpenAI embeddings: text-embedding-3-small is the low-cost text default at $0.02 per 1M tokens, a quarter of Embed 5 Fast’s $0.08, and text-embedding-3-large costs $0.13. Both are text only, so image and PDF-page search needs a multimodal model such as Embed 5.
• Voyage AI: now part of MongoDB, Voyage sells voyage-4 at $0.06 and voyage-4-lite at $0.02 per 1M tokens, with the first 200M tokens free across the voyage-4 family, and voyage-4-large matches Embed 5 Pro at $0.12. It belongs on any text-retrieval shortlist, and its rerank-3 matches Cohere’s reranker price rather than undercutting it.
• Google Gemini Embedding: gemini-embedding-2 adds audio and video to text and images, at $0.20 per 1M text tokens and $0.45 per 1M image tokens, while text-only gemini-embedding-001 costs $0.15. I pick gemini-embedding-2 when audio or video belongs in the index.
• Mistral embed: mistral-embed costs $0.10 per 1M tokens, between Embed 5 Fast and Pro, but takes text only with an 8K context against Embed 5’s 128K.
• Jina AI: now part of Elastic, Jina offers v5 text and omni embedding models, which makes it the natural shortlist entry for an Elastic-centered stack.
• Vectara and Ragie: managed RAG platforms that run ingestion, retrieval and answers as a service, where Cohere’s API hands over models and leaves the pipeline to me. Cohere’s closest managed answer is Compass, which is quote-only.

LlamaIndex and LangChain integrate Cohere’s embed and rerank models, but switching embedding vendors means re-embedding the whole corpus.

How I Reviewed Cohere

I ran Cohere’s two-stage retrieval flow on a trial key and a production key: Embed 5 Pro for documents and Embed 5 Fast for queries at several output dimensions with float, int8 and binary output, Rerank 5 Fast and Pro with Context Focus on, and Command A+ for grounded answers with citations and tool calls. I also ran OpenAI SDK code against the Compatibility API. The verdict weighs:

• Pricing and value: token rates against OpenAI, Voyage AI, Google and Mistral.
• Onboarding and limits: trial restrictions and the Go to Production workflow.
• Retrieval models: mixed inputs, compression, truncation and Context Focus.
• Generation and agents: citations, tool calls, North and model licenses.
• Deployment and security: channels, regions, retention and HIPAA scope.
• Alternatives: Jina AI, plus Vectara and Ragie for managed RAG.

Prices current as of November 2026.

Cohere Review: Should You Use Cohere for Search and RAG?

Yes, I recommend Cohere for enterprise search, RAG or agents over multilingual, visually rich or long documents. Among embedding models, it pairs the vectors with a same-vendor reranker and a private deployment path. Embed 5 costs $0.12 (Pro) and $0.08 (Fast) per 1M tokens, Rerank 5 adds $0.05 or $0.02, and Model Vault, a private VPC or an air-gapped site are there when data cannot go through the public API.

Start elsewhere if:

• Text cost rules: OpenAI’s text-embedding-3-small and Voyage’s voyage-4-lite cost $0.02 per 1M tokens.
• You want free volume: Voyage’s voyage-4 family includes the first 200M tokens free.
• Bedrock is mandatory: Embed 5 and Rerank 5 are not there.
• You want a managed service: Vectara or Ragie run ingestion, retrieval and answers for you.

Free Command A+ weights do not remove API limits: production keys stay at 1,000 calls a month on Command A+ until sales raises them. Older Command and Aya weights are CC-BY-NC, so self-hosting them commercially needs a Cohere agreement, and PHI stays off the SaaS API.

For a team ready to own its retrieval pipeline, I recommend Pro indexing, Fast queries and Rerank 5 Fast. Get a trial key at Cohere and build that pipeline on a noncommercial document set.

FAQ

Is Cohere free to use?

Yes, for prototyping. Every account gets a free trial key with 1,000 API calls a month, 20 chat requests and 10 rerank requests a minute, and no production or commercial use. Production keys are pay-as-you-go, billed at the end of each month or when the balance reaches $250, and Command A+ weights are a free Apache 2.0 download.

Is Cohere Embed better than OpenAI embeddings?

For images and PDF pages, yes: Embed 5 embeds text, images and fused pages in one vector across 100+ languages with a 128K context, while OpenAI’s text-embedding-3 models are text only. For plain text search, OpenAI’s text-embedding-3-small is cheaper at $0.02 per 1M tokens against Embed 5’s $0.08 to $0.12. See How Does Cohere Compare to Competitors? above.

Can I run Cohere models on-premises?

Yes. Private deployment runs in your VPC on AWS, Azure, GCP or OCI, on-premises or air-gapped, typically on Kubernetes, through Cohere sales. Model Vault is the managed middle ground at $3.00 to $12.50 per GPU-hour before promotional rates, with zero data retention by default and no rate limits.

Can I use Command A+ commercially?

Yes. Command A+ weights are Apache 2.0, so commercial self-hosting is allowed, and they fit on one B200 or two H100 GPUs. Command A, Command A Reasoning, Command A Vision, Command A Translate, Command R, R+, R7B and Aya weights are CC-BY-NC 4.0 and need a Cohere agreement for commercial use. On the API, Command A+ production limits go through sales.

How is Cohere Rerank priced?

Per token. Rerank 5 Pro costs $0.05 per 1M tokens and Fast costs $0.02. One request takes up to 10,000 documents with a 32,768-token context shared by the query and each document, and anything past 4,096 tokens per document is truncated by default. Production keys allow 1,000 rerank requests a minute, against 10 on a trial key.