API model string
command-r-08-2024 (pinned release; command-r alias also available — verify current alias behavior on cohere.com)Context window
128K tokens
Max output not published separately
- Knowledge cutoff
- not published
- Released
- August 2024 (08-2024 release; GA)
- Modalities
- text in; text out (RAG-optimized with citations)
Last verified July 10, 2026
Rate limits by tier
Cohere uses two key tiers — Trial and Production — with per-endpoint per-minute limits rather than a single shared pool. Trial keys are for evaluation only and capped at 1,000 total API calls/month across all endpoints. Production keys require an application form at cohere.com and support pay-as-you-go billing.
| Tier | Requirements | RPM | TPM | RPD | Concurrent | Notes |
|---|---|---|---|---|---|---|
| Trial key (evaluation only) | Free; no payment required. Not for production use. | 20 req/min (Chat endpoint) | not published | 1,000 API calls/month total across all endpoints | not published | Trial key global cap is 1,000 API calls/month total — not per endpoint. Once exhausted, all endpoints return 429 until the next calendar month. Newer Command A variants (Reasoning, Translate, Vision) behave like trial keys even on production plans — contact sales@cohere.com for production access. |
| Production key (pay-as-you-go) | Org account + application form on cohere.com. Pay-as-you-go billing; no monthly minimum. | 500 req/min (Chat endpoint) | not published | not published (unlimited for standard endpoints) | not published | Billing at month-end or when account balance hits $250. Rate-limit increases via support@cohere.com with use case description. |
| Per-endpoint limits (Trial → Production) | Limits vary by endpoint — not shared across endpoints | Chat: 20 → 500 | Rerank: 10 → 1,000 | Embed: 2,000 → 2,000 inputs/min (unchanged) | EmbedJob: 5 → 50 | Audio Transcriptions: 5 → contact sales | Tokenize: 100 → 2,000 | Default: 500 → 500 | not published per endpoint | not published per endpoint | not published | Limits are independent per endpoint per minute — hitting the Chat limit does not affect Rerank or Embed limits. Embed v4 handles PDFs natively. |
| Model Vault (enterprise dedicated instance) | Contact Cohere sales for pricing and access | Dedicated managed instance — not shared rate limits | not published (per-instance) | not published | not published | Per-instance pricing, not per-token self-serve. Required for Command A / flagship at production scale. |
Swipe the table sideways to see every limit column.
- 1.Trial keys are capped at 1,000 API calls/month total — not per endpoint. This is the most common cause of unexpected 429s on Cohere.
- 2.Rerank billing: 1 query + up to 100 documents = 1 search unit. Documents over 500 tokens are auto-chunked; each chunk counts as one document for billing.
- 3.No automatic prompt caching — implement app-level caching (Redis, application layer) to avoid redundant calls.
- 4.Embed v4 handles PDFs natively without pre-extraction. EmbedJob is the async batch alternative for large embedding workloads.
Limits verified against the Cohere docs, July 10, 2026.
Token pricing
What you pay per million tokens (USD). Input and output are billed separately.
Input
$0.15
per 1M tokens
Output
$0.60
per 1M tokens
- No automatic prompt caching — implement app-level caching in Redis or your application layer to reduce redundant token costs.
- Batch discount not documented for Command R self-serve.
- No long-context surcharge documented for Command R 128K window.
- Rerank billed per search (not per token): 1 query + up to 100 docs = 1 search unit; docs over 500 tokens are auto-chunked, each chunk counted separately.
- Embed billed per input (not per token).
- Command R family pricing for context: Command R7B at $0.0375/$0.15; Command R (this page) at $0.15/$0.60; Command R+ at $2.50/$10.00; Command A rates are sales-gated and not published for self-serve.
- Pin to command-r-08-2024 to avoid alias drift — the command-r alias may resolve to different versions with different pricing.
Side-project chatbot
$1.35
per month
Assumptions
5M input tokens / 1M output tokens per month (no caching available)
5 × $0.15 + 1 × $0.60 = $0.75 + $0.60 = $1.35/month
Mid-scale RAG pipeline
$16.20
per month
Assumptions
60M input / 12M output per month (no automatic caching — implement app-level)
60 × $0.15 + 12 × $0.60 = $9.00 + $7.20 = $16.20/month
High-throughput agentic system
$108
per month
Assumptions
400M input / 80M output per month
400 × $0.15 + 80 × $0.60 = $60 + $48 = $108/month
Run your own numbers
Drag your real monthly token volumes and watch the bill update live — priced against rival models at the same usage.
Estimated command-r-08-2024 (pinned release; command-r alias also available — verify current alias behavior on cohere.com) spend
$7.50/mo
Input: $4.50
Output: $3.00
30M in × $0.150 + 5M out × $0.600 = $7.50
Same volume, priced across models
- Mistral LargeCheapest$4.50
- command-r-08-2024 (pinned release; command-r alias also available — verify current alias behavior on cohere.com)This model$7.50
- Qwen3-Max$66
- GPT-5$300
Rivals priced at their published input/output rates for the same monthly volumes. Prompt caching is model-specific, so it is applied to command-r-08-2024 (pinned release; command-r alias also available — verify current alias behavior on cohere.com) only. Estimates for comparison; real bills vary with request shape and long-context surcharges.
command-r-08-2024 (pinned release; command-r alias also available — verify current alias behavior on cohere.com) vs the alternatives
Command R (08-2024) vs GPT-5.5, Mistral Small 3.1, and Qwen3-Max — contrasted on the metrics that matter for RAG and document-search use cases.
| Aspect | command-r-08-2024 (pinned release; command-r alias also available — verify current alias behavior on cohere.com) | GPT-5 | Mistral Large | Qwen3-Max |
|---|---|---|---|---|
| Context window | 128K tokens | ~1M tokens (GPT-5.5) | 128K tokens (Mistral Small 3.1) | 262K tokens (Qwen3-Max) |
| Input $/MTok | $0.15 | $5.00 (GPT-5.5) | ~$0.10–0.20 per third-party trackers — confirm in your console (Mistral Small 3.1) | $1.20 (Qwen3-Max) |
| Output $/MTok | $0.60 | $30.00 (GPT-5.5) | ~$0.30–0.60 per third-party trackers — confirm in your console (Mistral Small 3.1) | $6.00 (Qwen3-Max) |
| Prompt caching | no (app-level only) | yes (90% off) (GPT-5.5) | yes (Mistral Small 3.1) | not published (Qwen3-Max) |
| Native RAG stack | yes (Embed + Rerank + Command R) | no (GPT-5.5) | no (Mistral Small 3.1) | no (Qwen3-Max) |
| Batch discount | not published | 50% (GPT-5.5) | 50% (Mistral Small 3.1) | 50% (Qwen3-Max) |
| Free/trial tier | 1,000 calls/month (trial key, evaluation only) | no (GPT-5.5) | yes — experimental (Mistral Small 3.1) | 1M trial/model (Qwen3-Max) |
| Production key access | application form required | pay-as-you-go instant (GPT-5.5) | pay-as-you-go instant (Mistral Small 3.1) | pay-as-you-go instant (Qwen3-Max) |
Swipe the table sideways to see every model.
Hitting a 429? The playbook
The exact errors you'll see
HTTP 429 Too Many RequestsCohere returns HTTP 429 on exceeding per-endpoint per-minute limitsTrial key global cap: 1,000 calls/month total — triggers HTTP 429 on all endpoints once exhaustedWhy it happens & how to fix it
Trial key 1,000 calls/month global cap exhausted
Apply for a production key at cohere.com (application form required). The trial key cap is global across all Cohere endpoints — once hit, every endpoint returns 429 until the next calendar month regardless of per-minute limits.
Chat endpoint 500 RPM exceeded (production key)
Implement client-side rate limiting upstream of the API call — queue requests and release at 490 RPM max to leave headroom. No documented burst window. Use exponential backoff with the Retry-After header if present.
Rerank endpoint 1,000 RPM exceeded
Batch multiple queries into fewer Rerank calls where possible. Reduce document count per request. Pre-chunk documents to control auto-chunking billing behavior.
Embed endpoint 2,000 inputs/min hit
Queue embedding jobs. Use EmbedJob for large async batches — it bypasses the 2,000 inputs/min real-time limit. The EmbedJob limit is 50 jobs/min (production).
Command A endpoint returns 429 or trial-key behavior even on production plan
Command A and newer variants (Reasoning, Translate, Vision) behave like trial keys even on production plans. Contact sales@cohere.com for dedicated production access to these endpoints.
Retry strategy
Cohere returns HTTP 429 with a standard Retry-After header on rate-limit hits — honor the header value before retrying. For trial key monthly cap exhaustion (1,000 calls/month), no retry will succeed until the next month — upgrade to a production key instead. Use the cohere npm package or REST client; base URL is https://api.cohere.com/v1 with Bearer token auth. Implement exponential backoff: wait 2^attempt seconds with ±20% jitter. Contact support@cohere.com for production key rate-limit increases.
1import Cohere from 'cohere-ai';23const client = new Cohere.CohereClient({4 token: process.env.COHERE_API_KEY,5});67async function chatWithRetry(8 message: string,9 maxRetries = 510): Promise<string> {11 for (let attempt = 0; attempt < maxRetries; attempt++) {12 try {13 const response = await client.chat({14 model: 'command-r-08-2024',15 message,16 });17 return response.text ?? '';18 } catch (err: unknown) {19 const isRateLimit =20 err instanceof Error &&21 'status' in (err as Record<string, unknown>) &&22 (err as Record<string, unknown>).status === 429;23 if (isRateLimit && attempt < maxRetries - 1) {24 const jitter = 1 + Math.random() * 0.2;25 const wait = Math.pow(2, attempt) * 1000 * jitter;26 console.warn(`Rate limit — waiting ${Math.round(wait)}ms (attempt ${attempt + 1})`);27 await new Promise((r) => setTimeout(r, wait));28 } else {29 throw err;30 }31 }32 }33 throw new Error('Max retries exceeded');34}How to raise your limits
The ladder from the starter tier to enterprise — what each rung takes, and what it unlocks.
Trial key (evaluation)
Immediate.Sign up at cohere.com — trial key is issued immediately, free, no credit card required.
Unlocks: 20 RPM on Chat, 1,000 API calls/month total across all endpoints. For evaluation only — not production.
Production key (pay-as-you-go)
Review by Cohere team — typically days.Complete the production key application form on cohere.com (org account required). Pay-as-you-go, no monthly minimum; billing at month-end or when balance hits $250.
Unlocks: 500 RPM Chat, 1,000 RPM Rerank, 2,000 inputs/min Embed, 50 EmbedJobs/min. No monthly cap.
Rate-limit increase (production)
Days to weeks — reviewed by Cohere support team.Email support@cohere.com with your use case, expected volume, and current production key details.
Unlocks: Higher per-endpoint RPM beyond the default production limits.
Command A / flagship access
Custom negotiation.Contact sales@cohere.com — Command A rates are not published for self-serve. Describe your use case and scale.
Unlocks: Command A endpoint at production scale; per-token rates negotiated individually or via Model Vault dedicated instance.
Enterprise / Model Vault
Custom negotiation — typically weeks.Contact Cohere sales for a dedicated managed instance.
Unlocks: Dedicated instance with no shared rate limits, custom SLAs, and enterprise compliance features.
Cut your token spend
Use Command R7B for routing and classification
4× cheaper than Command R ($0.0375/$0.15 vs $0.15/$0.60 per MTok)Route simple classification, intent detection, and extraction tasks to Command R7B at $0.0375 input / $0.15 output per MTok. Escalate to Command R only for nuanced generation, multi-step reasoning, or RAG responses requiring citation quality.
Implement app-level caching
Eliminates redundant API calls — can reduce costs by 50–90% for repeated-context workloadsCohere has no automatic prompt caching. Cache system prompts, retrieval results, and frequent query responses in Redis or your application layer. Key on a hash of the full prompt; set TTL based on how often your context changes.
Use Embed + Rerank + Command R as an integrated RAG stack
Often cheaper than GPT + OpenAI embeddings for document-search pipelinesCohere's Rerank v3.5 per-search billing (1 query + 100 docs = 1 unit) combined with Embed v4 (PDF-native) and Command R citations creates a cost-effective end-to-end RAG pipeline without third-party embedding providers.
Use EmbedJob for large batch embedding
Bypasses 2,000 inputs/min real-time Embed limit for large ingestion jobsSubmit async EmbedJob requests for embedding large document corpora. EmbedJob limit is 50 jobs/min (production) — suitable for nightly batch ingestion pipelines.
Pin to command-r-08-2024 model version
Prevents billing surprises from alias driftThe command-r alias may resolve to different model versions over time with different pricing. Pin to the explicit version string command-r-08-2024 in production to guarantee stable billing and output behavior.
Pre-chunk large documents before Rerank
Avoids unpredictable Rerank billing from auto-chunkingDocuments over 500 tokens are auto-chunked by Cohere, with each chunk counting as one document for billing. Pre-chunk to your preferred size to control Rerank cost accurately and ensure consistent billing.
Implement client-side rate limiting at 490 RPM
Prevents Chat endpoint 429 cascades in high-concurrency applicationsNo burst window is documented for Cohere's 500 RPM Chat limit. Add a bottleneck queue (e.g., bottleneck npm package or custom token-bucket) at 490 RPM upstream of the API client to stay within limits without 429 interruptions.
Frequently asked questions
How do I get a Cohere Command R production API key?
Production API key access for Cohere requires completing an application form on cohere.com — it is not instant like most other providers. Create an org account, then submit the production key request with your use case. The trial key (free, 1,000 calls/month total) is available immediately but is for evaluation only. Contact sales@cohere.com if you need Command A or flagship access.
How do I increase Cohere Command R rate limits?
Email support@cohere.com with your use case, expected volume, and current production key details. For the Chat endpoint, the default production limit is 500 RPM; for Rerank it is 1,000 RPM; Embed is 2,000 inputs/min (shared between trial and production). For Command A and enterprise limits, contact sales@cohere.com.
Is Cohere Command R API free?
A trial key is available for free with 1,000 total API calls/month across all endpoints — suitable for evaluation only. Production use requires a paid production key. Pay-as-you-go pricing starts at $0.15/MTok input and $0.60/MTok output for Command R (08-2024). There is no free production tier.
Does Cohere Command R support prompt caching?
No. Cohere does not offer automatic server-side prompt caching for Command R. You must implement app-level caching (e.g., Redis or application-layer cache keyed on prompt hash) to avoid redundant API calls. This is the most important cost optimization for high-concurrency or repeated-context workloads on Command R.
What is the difference between Command R and Command R+?
Command R (08-2024, this page) is $0.15 input / $0.60 output per MTok and is optimized for RAG workflows with Cohere's Embed + Rerank stack. Command R+ (08-2024) is $2.50 input / $10.00 output — approximately 17× more expensive — and is positioned for more complex multi-step reasoning and enterprise use cases. Use Command R for document retrieval and citation-heavy generation; upgrade to R+ for tasks requiring deeper reasoning.
What does HTTP 429 mean on the Cohere API?
HTTP 429 Too Many Requests has two distinct causes on Cohere: (1) you exceeded a per-endpoint per-minute limit (Chat: 500 RPM production, 20 RPM trial; Rerank: 1,000 RPM; Embed: 2,000 inputs/min); or (2) your trial key's 1,000 calls/month global cap is exhausted — in which case no retry will help until next month. Check the response body to identify which limit was hit. Implement exponential backoff and honor the Retry-After header.
How does Cohere Command R compare to Mistral Large for RAG use cases?
Command R at $0.15/$0.60/MTok is dramatically cheaper than Mistral Large on a per-token basis, and it comes with Cohere's native Embed + Rerank stack (Rerank v3.5, Embed v4 with PDF support) — a full integrated RAG pipeline in one provider. Mistral Large at $2.00/$6.00/MTok has a larger 128K (vs Command R's 128K) — same context — but no native reranking. However, Mistral offers a 50% batch discount and prompt caching that Cohere lacks. For cost-sensitive pure RAG, Command R wins on price; for reasoning-heavy tasks, Mistral Large is stronger.
Can RapidDev help architect a Cohere RAG pipeline with Embed, Rerank, and Command R?
Yes. RapidDev builds production RAG architectures using Cohere's full stack — Embed v4, Rerank v3.5, and Command R — including app-level caching layers, rate-limit management, and cost optimization. Book a free scoping call at rapidevelopers.com/contact.
We build AI apps that don't hit rate limits
- Retry, backoff & caching built in
- Multi-provider fallback routing
- Fixed price, you own the code
30-min call. No commitment.