NotCheap MAKE EVERY CREDIT COUNT
Analyze Your Costs

Reka, Cohere or Mistral: API Bills and Private Deployment Costs

At 1.2 billion monthly tokens, the selected public API models range from $140.00 to $4,500.00 before extras. A private deployment needs a separate capacity and contract budget; token rates alone cannot tell you its cost.

Share this guide

Choose a deployable model before comparing its meter

A vendor name is not a SKU. The current Reka developer catalog lists Reka Flash 3 at $0.10 input and $0.20 output per million tokens. Our earlier Core and Flash figures are not the current catalog comparison. Flash 3 is a text reasoning model with a 64K context window; the supported-model table does not list native tool calling for it. That can rule it out of an agent workflow even when its bill looks attractive.

Cohere’s Command R documentation prices the dated command-r-08-2024 model at $0.15 input and $0.60 output per million tokens, with 128K context. Use the dated identifier: the unsuffixed command-r alias is deprecated in the model catalog. For a broader current enterprise option, Command A uses command-a-03-2025, 256K context and $2.50/$10 rates.

Mistral Small 4 is mistral-small-2603, with 256K context and $0.15/$0.60 rates. Mistral Large 4 is mistral-large-4, a public-preview model with 1M context and $1.36/$4.18 rates. Preview status matters to an enterprise rollout; price alone does not establish its support or stability commitments.

These rows are candidate budgets, not a benchmark claiming equivalent answers. Test the same documents, output format, languages, retrieval quality and tool requirements. Keep failed answers and human corrections in your acceptance record. A cheaper model that doubles review time can lose the comparison.

Two workloads, with input and output kept separate

Workload A is 100 million input plus 20 million output tokens: 120 million total tokens per month. Workload B is ten times larger: 1 billion input plus 200 million output, or 1.2 billion total. Both use a 5:1 input-to-output ratio. All figures below are USD public API usage before tax, discounts and ancillary services.

Same token quantities, different models; annual column is not a discounted contract
Model / API identifierInput / output per 1MA: 120M total / monthB: 1.2B total / monthB: twelve-month usage
Reka Flash 3
reka-flash-3
$0.10 / $0.20$14.00$140.00$1,680.00
Cohere Command R
command-r-08-2024
$0.15 / $0.60$27.00$270.00$3,240.00
Mistral Small 4
mistral-small-2603
$0.15 / $0.60$27.00$270.00$3,240.00
Mistral Large 4
mistral-large-4
$1.36 / $4.18$219.60$2,196.00$26,352.00
Cohere Command A
command-a-03-2025
$2.50 / $10.00$450.00$4,500.00$54,000.00

For Small 4, Workload B is 1,000 × $0.15 + 200 × $0.60 = $270.00. For Large 4 it is 1,000 × $1.36 + 200 × $4.18 = $2,196.00. Every multiplier is in millions of tokens. Multiplying the monthly result by twelve estimates a steady run rate; it does not mean the provider requires a year’s payment.

This table excludes embeddings, reranking, search tools, document parsing, storage and network charges. Use the enterprise RAG guide when those stages belong to the application. A cloud marketplace may have a different rate card from the vendor’s direct endpoint.

Output length can change the decision

Hold the workload at 1.2 billion total tokens and change only the mix. A report-writing application with 600 million input and 600 million output tokens costs more than a retrieval workload dominated by input. Do not carry the 5:1 result into a different application unchanged.

1.2 billion total tokens in both columns
Model5:1 input/output1:1 input/output
Reka Flash 3$140.00$180.00
Cohere Command R$270.00$450.00
Mistral Small 4$270.00$450.00
Mistral Large 4$2,196.00$3,324.00
Cohere Command A$4,500.00$7,500.00

Compare the provider-reported usage from representative requests. Separate reasoning or other billable output where the selected model counts it, rather than estimating cost only from the final visible answer. Pin the model version and token accounting used for the pilot.

Caching and batch processing are different savings

Large 4 lists cached input at $0.14 per million, compared with $1.36 for uncached input. If exactly half of Workload B’s billion input tokens qualify for that rate, the estimate is 500 × $1.36 + 500 × $0.14 + 200 × $4.18 = $1,586.00 per month. That is an assumed cache-hit workload, not a promise that half your prompts qualify.

Mistral’s batch documentation offers a 50% discount for asynchronous batch work, and the Large 4 model page lists batching support. Applying that discount to the uncached $2,196.00 example gives $1,098.00. Confirm endpoint and account eligibility; the job must tolerate asynchronous execution. We do not stack batch and cached-input discounts in these figures.

A general claim such as “up to 90% cheaper input” is not a rate for every model or every request. Use the exact model’s cached price and your observed hit fraction. Ask each vendor for committed-use terms separately; no unquoted enterprise discount is included above.

A private deployment needs a capacity quote, not a token shortcut

Decide what “private” means first: dedicated hosted capacity, inference inside your cloud account, or an on-premises system you operate. Cohere’s pricing FAQ treats VPC and on-premises arrangements as custom solutions. An API bill does not establish the price of that contract, nor does a private contract necessarily remove every token-based fee.

There is no universal minimum of eight H100 GPUs. Cohere documents Command A running on two A100 or H100 GPUs; that is a model-serving footprint, not a complete high-availability service. The number of replicas and GPUs still depends on precision, context, concurrency, throughput and failover requirements. Benchmark the configuration you intend to buy.

Licensing also varies by exact weights. Small 4’s model page specifies Apache 2.0; that does not make its hosting free. Reka publishes Edge weights, but its Edge license contains separate production and hosted-service restrictions. Obtain terms for the intended deployment rather than assuming a public model download grants every commercial use. Hosted Flash 3 pricing is not an Edge self-hosting license.

Build a monthly private budget as: vendor licence + GPU capacity + CPU/RAM + storage/transfer + support + operations time + setup cost allocated over the chosen period. Add any contractual usage fees. Include idle capacity and the second replica if the availability design needs one. Request the architecture and price together.

Budget illustration only: these allowances do not specify a tested GPU configuration
Illustrative planning allowanceMonthly amount
Compute allocation, not a vendor quote$2,500.00
Storage and transfer allowance$100.00
Support allowance$200.00
Operations: assumed 10 hours × $100$1,000.00
Recurring subtotal, before vendor licence$3,800.00
One-time setup: assumed 40 hours × $100$4,000.00
First month, including setup, before licence$7,800.00

This $3,800 allowance is a screening exercise, not a quote or a capacity claim. At Workload B it exceeds the Small 4 public API estimate by $3,530 and the Large 4 estimate by $1,604 before licensing. It is $700 below Command A’s API estimate, but that does not prove it can serve Command A’s workload. Capacity, quality and contract terms must all match before calling anything a saving.

At zero API usage the token component is zero, although balances or contracted commitments are separate. A private service that stays provisioned continues to incur its fixed costs. There is no defensible universal “10 billion tokens” break-even point without measured throughput and an actual deployment quote.

What to ask for before signing

  • Accepted workload: model version, languages, context length, tool use and quality threshold.
  • Capacity: sustained output tokens per second, concurrency, latency targets, peak load and a failover test.
  • Data terms: region, retention, training use, logs and who can administer the environment.
  • Complete price: licence, infrastructure, support, minimum commitments, overages, renewal and migration costs.
  • Lifecycle: pinned versions, deprecation notice, preview limitations and the cost of upgrading.

Start with the public API pilot if it satisfies the deployment requirements, measure accepted results, and request private quotes only for a defined workload. A private environment may be worth paying for when control or deployment location is a hard requirement. Treat that as a purchasing requirement rather than inventing token savings to justify it.

For a broader public-model shortlist, see AI API costs by workload. For the larger model’s preview and caching details, see our Mistral Large 4 pricing guide. Report a correction with the affected model and official source.

PASS IT ON

Useful for your team?

Share this guide with a colleague.