Break down input vs output pricing, cache hits, batch discounts, monthly projections, and the workflow to estimate prompt cost before you ship.
Try it in TokenCALC
Set output size, cache hit %, and users × messages/day to project monthly spend.
Prompt cost starts with tokens, not words
Prompt cost is the money you spend for the tokens in a request and its completion. Providers publish USD per million input tokens and per million output tokens. Until you have token counts, you do not have a real cost model.
Convert words to tokens only for early ballparks. Then paste the real prompt into TokenCALC.
Input and output are priced separately
Your request cost is roughly (input_tokens × input_rate + output_tokens × output_rate) / 1,000,000. Output is usually more expensive. A tiny prompt with a long answer can cost more than a long prompt with a short answer.
Reasoning-style models may bill internal thinking at the output rate. Read model notes when you project those workloads.
Cache and batch change effective rates
Prompt caching discounts repeated prefixes when published. Batch APIs often cut standard rates with higher latency. TokenCALC only enables these controls when the catalog has rates for that model.
For deeper cache math, use the prompt caching guide and the cache savings tool. For batch, use the batch pricing tool, then confirm on the main cost calculator.
From one request to a monthly budget
Multiply per-request cost by users and messages per day to estimate daily, monthly, and yearly spend. This is where product traffic assumptions matter more than a single list price.
Compare models on the same workload before you commit. A slightly higher input rate can still win if output is shorter or cache hits are higher.
Worked thinking example (structure only)
Suppose a support bot sends about 800 input tokens and expects about 300 output tokens per ticket. Monthly tickets are users × messages per day × days. Plug those token counts and volume into the cost calculator rather than inventing arithmetic in a vacuum.
Do not copy invented USD totals from blog posts. Use curated model rates with lastVerified dates, then confirm on the provider page for critical budgets.
Exact vs Approx in cost estimates
OpenAI Exact counts make OpenAI-side estimates stronger. Approx counts for other providers need margin. Count accuracy and price accuracy are separate; you need both.
See Exact vs Approx and the provider bridge guides when the label changes your confidence.
Common mistakes
Avoid these prompt-cost errors.
- Pricing from word counts
- Ignoring output tokens
- Forgetting cache write costs
- Assuming batch and real-time traffic share one SLA
- Comparing models without fixing the same prompt and volume
Real-world scenarios
A summarization feature sends 2,000 input tokens but requests 800 output tokens. Finance priced only the input side using a word count heuristic and underestimated monthly spend by a wide margin because output rates are higher.
A marketing chatbot reuses a 1,500 token brand guideline block on every message. After enabling prompt caching in production, input spend drops on cache hits, but the first request in each session still pays full input plus any published cache write fee.
A batch nightly job processes support tickets with identical volume to the daytime realtime bot, but qualifies for batch rates. Mixing both traffic streams in one spreadsheet without separating SLAs produces meaningless averages.
Step-by-step in TokenCALC
Count input tokens on the tokenizer or main calculator with your representative prompt. Set expected output tokens explicitly: short replies, long reports, and JSON payloads differ dramatically in cost.
Open the cost calculator doorway. Pick the model, enter or confirm token counts, set cache hit percentage when the catalog publishes cache rates, and enter users times messages per day.
Review daily, monthly, and yearly projections. Switch models without changing the prompt to compare apples to apples. Export mental notes for finance: Exact vs Approx label, lastVerified rate date, and assumptions about output length.
Related concepts
Prompt caching and batch pricing are modifiers on the base input rate. LLM API pricing covers the full billing picture. Exact vs Approx affects how much trust to place in the token side of the equation.
The cache savings and batch pricing tools isolate those levers before you return to the main calculator for combined projections.
Expert notes
Per request cost times volume is simple arithmetic, but volume assumptions dominate uncertainty. Pilot with real user behavior before you multiply a demo prompt by a million seats.
Some products charge end users per seat while COGS scales per token. Unit economics need both sides: your margin per active user and tokens per active user per day.
Revisit forecasts when you change default models, max output limits, or retrieval chunk sizes. Prompt cost is not a one time spreadsheet cell.
Warnings
Do not copy USD totals from third party blog posts with invented numbers. Use curated catalog rates with lastVerified stamps and confirm on the provider page for board level budgets.
Free tier credits and promotional pricing are not modeled in TokenCALC. Subtract credits manually when comparing to invoices.
Frequently asked questions
How is prompt cost calculated?
Multiply input and output tokens by per-million rates, then scale by request volume. Apply cache or batch only when those rates exist for the model in the catalog.
Why is my short prompt expensive?
Long outputs or reasoning tokens can dominate. Output rates are usually higher than input rates, so completion length matters as much as prompt length.
Can I project monthly prompt cost?
Yes. Use users times messages per day in the cost calculator after you have per-request tokens for a representative workload.
Do cache hits reduce prompt cost?
They can, when a stable prefix repeats and the provider publishes cached-input pricing. Output tokens are unchanged by cache hits.
Where do I start in TokenCALC?
Count tokens on the tokenizer or main calculator, then switch emphasis to the cost calculator doorway for spend and volume.
Should input and output use the same token count assumptions?
No. Measure or estimate them separately. A concise prompt with a verbose answer is a common and expensive pattern.
Next steps
Use the calculator links above for Exact or Approx counts on your own prompts, then browse related guides and model pages to compare pricing assumptions.