What will your LLM
application actually cost?
Estimate monthly and annual LLM inference cost using the workload variables that matter in production: request volume, token usage, model calls per workflow, retry rate, caching, model routing, and growth.
See cost per completed task, a 12-month projection, sensitivity at different traffic levels, and which optimization lever is likely to save the most. No account is required, and calculations run in your browser.
Your workload
Model routing split
Estimated cost
Sensitivity — monthly cost at different volumes
| Daily volume | Monthly cost |
|---|
Why simple token estimates
understate real cost.
A basic AI token calculator multiplies input and output tokens by a model's published price. Useful for one isolated API call — but most production AI features involve more than one call.
A single user request may trigger planning, retrieval, classification, tool use, validation, synthesis, and retries. Some requests are served from cache. Others route to different model classes. Traffic may grow faster than the initial forecast.
The number that matters to a product team isn't only cost per model call — it's cost per completed user task at expected production volume. This calculator models the main multipliers around the model call so teams can make a more defensible estimate before launch.
What this calculator
estimates.
Combines daily request volume, average input/output tokens, model calls per workflow, retry rate, cache-hit rate, routing split, published prices, and expected growth.
Returns: estimated monthly and annual inference cost, cost per completed task, billed request count, a 12-month projection, sensitivity at different traffic levels, and the estimated impact of caching, routing, and fewer retries.
How the
calculation works.
Five steps from raw traffic to a monthly number.
Billed requests = daily requests × 30 × (1 − cache-hit rate)Total model calls = billed requests × calls per workflow × (1 + retry rate)See our LLM inference cost optimization article for the reasoning behind each lever.
The variables with
the largest impact.
Not every input matters equally. These six have the largest effect on the final estimate.
Traffic is the base multiplier. Use expected production traffic, not current prototype or internal-test volume. If demand is uncertain, use the sensitivity table to compare conservative, expected, and high-growth cases.
Long system prompts, large retrieved contexts, conversation history, and verbose outputs all increase cost. Measure token usage from representative traces when possible, rather than relying only on prompt length estimates.
Frequently overlooked. An agent may call a model to plan, select a tool, interpret the result, validate output, and generate the final response — a workflow with five cheap calls can cost more than one larger-model call.
Retries come from timeouts, rate limits, malformed structured output, tool errors, or failed validation. Set explicit retry limits and monitor the reason for each retry rather than treating retries as free resilience.
Caching prevents repeat or sufficiently similar requests from reaching the model. The achievable rate depends on the use case, freshness requirements, personalization, and whether semantic caching is appropriate.
Routing simpler tasks to a smaller model can cut cost substantially, but only if quality is measured by task. Classification, extraction, and formatting may tolerate a smaller model; complex reasoning may not.
What the estimate
includes — and doesn't.
The result estimates model-inference cost based on the values entered. It's not a complete total cost of ownership calculation. Depending on the architecture, a full budget may also need:
- Embedding generation and re-indexing
- Vector database or search infrastructure
- Reranking models
- Speech-to-text and text-to-speech
- Image or document processing
- Agent tools and third-party API charges
- Evaluation and model-judge calls
- Observability and trace storage
- Compute, queues, databases, and network transfer
- Engineering, support, and human-review time
Published model pricing also changes — confirm current provider pricing before approving a budget.
How to reduce LLM
cost safely.
Cost reduction that doesn't quietly trade away quality.
Cheap calls aren't useful if they produce more failures, escalations, or retries. Track cost alongside task completion and quality.
Use smaller models for bounded tasks that meet a measured quality threshold. Reserve more capable models for requests that require them.
Retrieve only the information needed for the current task. Remove duplicated instructions and avoid carrying unnecessary conversation history indefinitely.
Cache stable results, deterministic transformations, and repeated retrieval where the product's freshness and personalization requirements allow it.
Set maximum attempts, steps, time, and spend per workflow. Alert when actual behavior approaches those limits.
Where latency isn't user-facing, batching can reduce provider or infrastructure costs. Confirm the provider's current pricing and service constraints.
When to run
this calculator.
Moments where a defensible number actually changes a decision.
- Before approving an LLM feature or AI agent budget.
- When comparing two models or routing strategies.
- Before moving from an internal pilot to customer traffic.
- When an agent uses multiple calls per task.
- After an unexpected increase in an AI provider invoice.
- During an architecture review or capacity-planning exercise.
- When deciding whether caching or smaller-model routing is worth implementing.
Before you take this
to finance.
What this estimate is, and its limits.
How accurate is this LLM cost calculator?
The estimate is only as accurate as the workload assumptions entered. The strongest inputs come from representative production or load-test traces. For a new system, calculate a range using conservative, expected, and high-usage scenarios. Pricing shown is illustrative, last checked 2026-10-14.
What is the difference between a request and a model call?
A request is the user-visible task entering the application. A model call is one invocation of an LLM API. One request can produce several model calls when a workflow plans, retrieves data, calls tools, validates output, retries, or synthesizes a final answer.
Does the calculator work for AI agents?
Yes. Enter the average number of model calls used to complete an agent task and include the observed or expected retry rate — see our AI Agent Development page. For agents with highly variable paths, calculate separate estimates for common, complex, and failure scenarios.
Why are input and output token prices different?
Model providers commonly price input and generated output tokens separately. Output tokens are often more expensive because generation requires sequential inference. Always confirm current prices with the provider.
What cache-hit rate should I use?
Use an observed rate if the application already has caching. Otherwise, start with zero for the baseline and model alternative rates as optimization scenarios. Don't assume a high cache-hit rate for personalized, rapidly changing, or permission-sensitive content.
Does this include RAG infrastructure cost?
Not by default. The calculation focuses on model inference. Embeddings, vector storage, search, reranking, document processing, and data pipelines may need to be budgeted separately — see our RAG Development page.
Does Dwayo store the values entered?
The calculation runs in the browser. Values are only shared with Dwayo if you explicitly choose to save an estimate or request an engineering review. See our privacy policy.
A useful estimate depends on
accurate architecture assumptions.
Build the estimate using your expected workload. If calls per workflow, retry behavior, routing, or token usage are unclear, a Dwayo engineer can review the assumptions against the intended architecture.