LLM Token Cost Calculator
Estimate what an AI feature could cost each month, and compare models to find lower-cost options to evaluate.
Workload example
Illustrative starting points, not measurements of your usage.
40 conversations a day, 5 model calls each: 200 calls. Each call sends about 1,500 tokens (instructions, earlier messages, the new message) and gets about 300 back. No caching or batch.
Your workload
Tokens are the units providers use to measure model input and output. Each provider counts the same text in its own way.
Count every call to the model, not just users or conversations. One conversation may make several calls.
Average per calendar day, weekends included: about 6.1K calls in an average month. Counted working days only? Multiply by 5 and divide by 7: 200 becomes about 143.
Everything sent to the model: instructions, the new message, earlier messages, retrieved content, and tool information where applicable.
The tokens generated for each call. Use the provider's reported usage when available; visible answer length may not include every billed token.
Estimate from sample text (optional)
Do not paste credentials, personal information or confidential customer content. The estimate is worked out in this browser tab; the text is not sent anywhere or saved in the link.
Everything one call sends: the instructions and context normally sent, not only the customer's latest question. In a conversation, later calls usually send the earlier messages again.
A typical response the model gives back. Input text is never used to guess the output.
About 4 characters per token for English text. Each provider counts differently, and code, JSON and other languages vary more. Sample text cannot show hidden reasoning tokens, image or audio units, or tool charges: use the provider's reported usage for those.
Models to compare
Pick up to 6. Prices are list prices for the same token counts.
Optional: more assumptions
Prompt cachingOff
Batch processingOff
Cost per conversation or task5 calls each
Sets the cost per conversation or task. Model API calls per day stays as entered.
Compare the cost per conversation or task with what each one earns or saves.
Agent what-ifOff
Multiplies model calls by 2.4 to show how an agent loop could raise the cost. 2.4 is an illustrative assumption, not a measurement. If your call count is measured, leave this off: it already includes the loop.
Monthly cost
Example: Support chat$1.83 to $36.53 a month
Across 4 models, at 200 model API calls a day, 1,500 input and 300 output tokens per call, in a 30.44-day average month.
These are example numbers, not your usage. Change any input to make this your own estimate.
| Model | Monthly cost | Per call | Per conversation | Next 12 months | vs lowest |
|---|---|---|---|---|---|
GPT-6 Luna $0.100 in / $0.500 out per 1M SmallLowest estimate Price notePrompts over 272K tokens bill the whole request at 2x input and cache rates and 1.5x output ($0.20 in / $0.75 out). | $1.83$0.0015 per conversation | $0.00030 | $0.0015 | $22 | Lowest |
Gemini 3.8 Flash $0.750 in / $3.75 out per 1M WorkhorsePrice rises 1 Jan 2027 Price notePrice runs through 31 Dec 2026; from 1 Jan 2027 it doubles to $1.50 in / $7.50 out (cache read $0.15). Cache storage, $0.50 per 1M tokens per hour through 2026 and $1.00 after, is not modeled. | $13.70$0.0113 per conversation | $0.00225 | $0.0113 | $302 | +$11.87 |
Claude Sonnet 5.5 $2.00 in / $10.00 out per 1M Workhorse Price note1M-token context window at standard pricing. | $36.53$0.0300 per conversation | $0.00600 | $0.0300 | $438 | +$34.70 |
GPT-6.1 Sol $2.00 in / $10.00 out per 1M Workhorse Price notePrompts over 272K tokens bill the whole request at 2x input and cache rates and 1.5x output ($4 in / $15 out). | $36.53$0.0300 per conversation | $0.00600 | $0.0300 | $438 | +$34.70 |
The same text can count as more tokens on some models (Claude 4.7 and later count about 30% more). This table prices the same token count on every model.
A lower token price does not mean the same quality or speed: test quality, latency and token usage before choosing a model. Next 12 months is today's volume with published price changes from their dates, with no growth.
A lower-cost model to evaluate. At the same entered token counts, GPT-6 Luna would cost $34.70/month (95%) less than GPT-6.1 Sol. Potential cost difference only: test quality, latency, and token usage before switching.
Where the money goes
Each model's monthly cost split into input, cached input, cache writes and output, at these token counts and that model's rates.
Growth scenarioOff
A what-if, not a forecast: enter the monthly growth you expect and see the three lowest estimates over 12 months. 0% shows today's volume held flat.
Prices verified 30 Oct 2026. Public list prices in USD, before any discount you have negotiated.
Want your real AI spend, not an estimate?
Finitizer Tokenomics: Reads your AI bills and exports and shows what changed, who spent it, and what to fix. No proxy, no code change. Founding customers list now open.
How this calculator works
Everything runs in your browser against a hand-checked price list. Five assumptions are worth knowing:
- Writing to the cache is charged. Some providers charge extra to cache a prompt. A shared prompt is written about once every five minutes while traffic flows; a conversation cached turn by turn writes every new token. If little of the input is reused, caching costs more than it saves, and the results show that.
- We use 30.44 days for an average planning month. Model API calls per day is an average over calendar days, weekends included.
- Every model is priced at the same token counts. A lower estimate is a price difference, not proof of equal quality or speed, and the same text can count as a different number of tokens on another model.
- Batch and caching are applied as independent shares of traffic. Providers differ on whether the two compose. Where a provider says they do not (Groq), batch requests pay the batch rate on their full input.
- Dated price changes count. Monthly cost is today’s price. Next 12 months and the growth scenario price each month at the rate in force, so a published rise shows up from its date. Next 12 months holds today’s volume flat; growth is a separate what-if.
Questions
How do I calculate LLM API costs?
Multiply your average input tokens per call by the model input rate and your average output tokens per call by the output rate (both quoted per million tokens), add them for a cost per call, then multiply by model calls per day and 30.44 days. Many models price output above input, so check both rates. This calculator does that for several models at once, with optional caching and batch rates.
Does prompt caching always save money?
No. Some providers charge extra to write a prompt into the cache, and cache lifetimes, write charges and read discounts vary by provider, model and API. If little of the input is read back from the cache, the writes can cost more than the reads save. Anthropic, and OpenAI from GPT-5.6 on, charge for cache writes; DeepSeek does not. The hit rate to enter is cached input tokens as a share of all input tokens. This calculator uses each model's listed cache rates and flags a model when caching costs more than it saves.
What is the batch API discount?
Several providers offer a lower rate for requests sent through an asynchronous batch endpoint, in return for a delayed result. The discount and the completion window vary by provider and model; this calculator uses each model's listed batch rate, and a model without one pays full price. It suits work that does not need an immediate response, such as summarization, enrichment, evaluation and offline classification.
Why do agentic workloads cost so much more?
An agent re-sends everything it has read so far on every step of its loop, and may retry failed steps, so one task can become many model calls with a growing prompt. The Agent what-if multiplies calls by 2.4, an illustrative assumption rather than a measurement, and is off unless you turn it on. If your call count is measured it already includes the loop. The AI Agent Cost Calculator prices a loop step by step.
What is cost per conversation and why does it matter?
Cost per conversation, per agent run, or per document is the unit economic that connects AI spend to business value. A total monthly figure cannot tell you whether a feature is profitable; a cost per unit compared against the revenue or saving per unit can.
Related tools
AI Agent Cost Calculator
Agents make many model calls per task. Price the whole loop, not one request.
OpenAI Usage Export Analyzer
Drop in your OpenAI or Claude usage export and see where the AI spend goes.
OpenAI Carbon Calculator
The carbon of AI prompts and GPUs, from the figures the makers have published.
OpenMore tools for: AI and model usage
This is an estimate. Finitizer shows your actual AI spend.
Finitizer Tokenomics reads what your AI actually costs from your own bills and exports (Anthropic, OpenAI, Google, AWS Bedrock, Azure OpenAI and more), ties it to the workloads and owners it serves, forecasts the months ahead with a range, and sets each workload's spend against the value it returns. Measure against what you spend, not what you estimated.
See Finitizer Tokenomics