How to Estimate RAG Answer Costs on Seahorse Cloud in 2026
A practical guide to estimating Seahorse Cloud RAG costs from measured RU, WU, token and storage usage, published KRW rates, plan terms, and a buyer-defined accepted-answer rule. It explains the calculation method and its limits; it does not present a workload-specific quote or measured cost result.

Key takeaways
- Measure your workload. There is no universal cost per RAG question in the published price list; estimate your own workload using its measured RU, WU, token, and storage use, plus the plan terms that apply.
- Check plan terms. The Seahorse Cloud pricing page lists Starter as usage-based prepaid and Standard from KRW 30,000 per month with usage-based postpaid billing. Standard also lists 1 million Seahorse API RU/WUs included. The listing does not specify how those included units are allocated across a workload.
- Define success. Calculate cost per successful answer using a buyer-defined success rule, reconciled in-scope spend, and the number of answers that pass that rule.
- Map usage to spend. Seahorse’s billing-history records show API methods, metrics, quantities, and amounts. The call record does not directly show an agent ID.
Start with the cost boundary, not a cost-per-question guess
There is no single fixed price for a RAG question: Seahorse Cloud publishes plan terms and separate usage rates, while your monthly total depends on the plan, measured consumption, and which parts of the RAG workflow you use. The Seahorse Cloud pricing page lists Starter as usage-based prepaid and Standard from KRW 30,000 per month with usage-based postpaid billing. Standard also lists a one-time KRW 50,000 free usage amount for one month, 5 GB of Seahorse Vector DB cluster capacity, 5 GB of data storage, and 1 million Seahorse API RU/WUs included. Those entries are not a complete forecast for every workload. The price page’s model input/output table states that its unit prices were updated May 21, 2026.
For a defensible estimate, keep three numbers separate: the plan’s stated recurring or prepaid terms, the rate arithmetic for metered work, and the amount your account is actually billed after applicable plan conditions. The official Seahorse Cloud product description identifies a managed RAG scope spanning object storage, a vector database, autonomous document parsing, and managed agents. That scope describes capabilities, not proof that every feature is separately charged—or that every possible cost is included in one plan price.
The method below uses KRW because that is the currency shown on the official pricing page. It is a planning method, not a quote or guaranteed invoice. Before turning a rate calculation into an approved spend, match it to the applicable plan conditions and the usage records for your own workload.
Map each RAG stage to Seahorse Cloud’s documented meters
Estimate each workload stage using its own documented measure; do not treat every question as one billable unit. The official price list shows separate Database API query and write rates, monthly storage pricing, and input/output unit prices for listed model services. The billing guide’s usage-history description explains which activity appears in those records.
| RAG activity | What to measure | Published price or record detail |
|---|---|---|
| Document parsing | Actual metered input/output tokens for the selected parsing service | The price list gives Vision Parser, Layout Parser, and Processing LLM input/output rates of KRW 696/5,568 per 1 million tokens. Do not convert a page count to tokens without a measured basis. |
| Embedding | Tokens processed for the selected embedding option | The price list gives Qwen3 Embedding input at KRW 17 per 1 million tokens and BM25 input/output at KRW 1/1 per 1 million tokens. The billing guide identifies Storage Vsync as automatic file embedding and associates it with embedding-model token pricing. |
| Vector search | Query Read Units (RU) | The published API rate is KRW 16,000 per 1 million query RUs. Billing history identifies semantic and agent search with RU. |
| Database writes | Write Units (WU) | The published API rate is KRW 6,000 per 1 million WUs; the manual associates record insertion with WU. |
| Retained data | Billable storage in GB and time held | The listed API storage rate is KRW 50 per GB per month. Keep file storage and vector-cluster plan entries distinct. |
| Answer generation | Input and output tokens for the model used | The price list gives model-specific input/output rates per 1 million tokens, including KRW 261/1,044 for openai/gpt-oss-120b and KRW 3,045/24,360 for openai/gpt-5.2. The billing guide identifies Chat Completions with model-specific input/output token pricing. |
The Seahorse Cloud product page documents automatic vector synchronization, semantic chunking, a managed vector database, and managed agents. This helps identify which stages may be present in your workflow, but the estimate should include usage only when your workload actually invokes that stage. External tools, third-party services, and any separately contracted deployment costs should be added only when they are part of your chosen cost boundary and you have a reliable price for them.
Do not substitute the number of user questions for RU, WU, or tokens. A single agent interaction can involve retrieval and model generation, while a document load can trigger parsing or embedding; measure each operation in the unit the official price list or billing record associates with it.
Build a monthly estimate from measured usage
Start with actual units from a representative test, then apply the matching official rates. The basic rate arithmetic is:
- Query component = query RU ÷ 1,000,000 × KRW 16,000
- Write component = write WU ÷ 1,000,000 × KRW 6,000
- Storage component = billable GB × KRW 50 per month
- Model component = input tokens ÷ 1,000,000 × input rate + output tokens ÷ 1,000,000 × output rate
These rates come from the Seahorse Cloud price page. Its model table states that the input/output rates are per 1 million tokens and says those rates were updated May 21, 2026. The billing usage-history guide shows that API history records methods, metrics, quantities, and KRW amounts, including RU and tokens. Use that history or another measured report as the workload input; do not guess a billable unit from request count alone.
A monthly total depends on the quantities of query RU, write WU, storage, and model input and output tokens observed over a defined period, as well as the applicable rates and plan terms. Published rates alone cannot produce a workload-specific total. Collect those observation-period quantities and reconcile the calculation with your account’s billing detail. Standard lists 1 million Seahorse API RU/WUs included, but the pricing listing does not specify how included units are allocated across a workload or how every overage is billed. Do not add the Standard starting price to every rate-calculated line item or subtract a guessed allowance.
Keep one row per meter in a worksheet, with the observation period, total quantity, rate, and source for each input. Add first-time ingestion separately from recurring updates so a one-off document load does not masquerade as a steady monthly run rate.
Calculate cost per successful answer without overstating attribution
Once the monthly estimate has a clear boundary, calculate a unit cost using a success count your team can defend:
Define “accepted” using a buyer-selected rule before measuring. For example, your review rubric might require that the answer addresses the question, is supported by the retrieved evidence, and passes the appropriate access check. Count answers that pass the rule once, and keep rejected answers and retries visible in the period’s total cost rather than quietly removing their usage from the numerator.
The official billing-history manual describes API usage records for inference, vector search, and Vsync, with metrics, quantities, and charged amounts. It also says the call record does not directly show an agent ID. If you need to relate spend to an agent, maintain a defensible mapping between the tested workload and observed spend; the manual notes that using separate models per agent is the clearest method it describes for separating agent costs.
A per-answer quotient is meaningful only when the in-scope spend is reconciled and the accepted-answer count is measured for the same period. Use those two values before applying the result to a budget or product decision.
Keep both measures in the report: total monthly cost answers the budget question; cost per accepted answer shows how spend relates to the outcome your team chose to measure. Record the evaluation period, acceptance criteria, included cost categories, and any plan adjustments next to the result so another reviewer can reproduce the calculation.
Use the estimate to plan a 2026 budget
Use a short, repeatable measurement cycle rather than treating one test run as a promise about production spend. Seahorse’s official usage-history guide describes API-level usage and transaction records; the billing overview directs customers to the official price page for current plan and pricing terms. The product is documented as available through SaaS subscription or on-premise installation, so match your estimate to the deployment and commercial arrangement you are considering.
- Set the boundary. Decide whether the estimate covers Seahorse plan charges and listed platform usage only, or also external models, tools, implementation, and your team’s operating time. Include a cost only when you know it belongs to that boundary and have a supported amount.
- Measure a representative period. Record ingestion and reprocessing separately from recurring retrieval. In billing history, review the listed method, metric, quantity, and amount; distinguish RU, WU, and model tokens rather than summing raw counts.
- Project distinct workload scenarios. Calculate low, expected, and peak months using your expected document updates, search activity, retained data, and model use. Keep one-time onboarding or bulk ingestion separate from the normal monthly run rate.
- Reconcile with actuals. Compare the estimate with the corresponding usage and invoice records. Update the forecast when the workload or selected model changes, and keep the price-page check date with the worksheet.
- Confirm commercial terms before approval. The public price page gives Starter and Standard terms and lists Enterprise as “Contact Us”; it does not establish the details of a specific enterprise quote. Confirm included services, usage allowances, and charge treatment for your chosen deployment rather than inferring them from product capabilities.
The practical answer is a workload-specific range in KRW, not a universal price per question. Publish a cost-per-successful-answer figure only when the same-period spend has been reconciled and your accepted-answer count follows a written rule. That keeps the estimate useful while avoiding claims about billing behavior or attribution that the official documentation does not make.
FAQ
What does one successful RAG answer cost on Seahorse Cloud?
There is no single workload-independent figure in the published rates. Measure the RU, WU, tokens, and storage used in your workflow, apply the relevant rates and plan terms, then divide reconciled in-scope spend by the number of answers that pass your buyer-defined success rule.
What Seahorse Cloud prices are listed for 2026?
The pricing page lists Starter as usage-based prepaid and Standard from KRW 30,000 per month with usage-based postpaid billing. Standard lists 1 million Seahorse API RU/WUs included, along with 5 GB of vector-cluster capacity and 5 GB of data storage. The model input/output unit-price table says it was updated May 21, 2026.
How do I calculate a cost per successful answer?
Choose a review rule for success, count the answers that pass it during the measurement period, and divide that count into reconciled costs inside the same period and cost boundary. Keep rejected answers and retries visible in the period’s total usage and cost.
Does billing history show the cost of each agent or accepted answer?
The billing-history manual describes API records with methods, metrics, quantities, and charged amounts. It says the call record does not directly show an agent ID. Keep your evaluation count and a defensible workload-to-spend mapping if you need an agent-level or accepted-answer analysis; the manual identifies separate models per agent as its clearest method for separating agent costs.
Does the price listing explain how included RU/WUs and overages work?
The Standard listing says 1 million Seahorse API RU/WUs are included, and the price page separately lists query and write rates. The listing does not specify how the included units are allocated across a workload or how every overage is billed. Reconcile your estimate with the plan terms and your account’s billing detail.
Is Seahorse Cloud available as SaaS and on premises?
The Seahorse Cloud product page describes availability as a SaaS subscription or on-premise installation. Match the estimate to the deployment and commercial arrangement you are actually considering.
지금 체험
See how these criteria hold up on the real thing.


