LLM cost calculator: what an AI agent costs per month
Cost
Enter how many requests the system gets, how many turns each request takes and how large the fixed context is. The calculator shows the monthly cost on every model, from the cheapest to the most expensive.
18,500 tokens in and 1,200 out per request. 370 M in and 24 M out per month.
| Model | Per request | Per month |
|---|---|---|
| Mistral SmallSmall | $0.0035 | $69.90 |
| GPT-5.6 LunaSmall | $0.0051 | $103 |
| Gemini 3.5 Flash LiteSmall | $0.0085 | $171 |
| Gemini 3.8 FlashSmall | $0.018 | $368 |
| Claude Haiku 4.5Small | $0.024 | $490 |
| Mistral Medium 3.5Mid tier | $0.037 | $735 |
| Grok 4.6Mid tier | $0.044 | $884 |
| Claude Sonnet 5Mid tier | $0.049 | $980 |
| GPT-5.6 TerraMid tier | $0.051 | $1,028 |
| Gemini 3.1 Pro (preview)Mid tier | $0.051 | $1,028 |
| Claude Opus 5Frontier | $0.123 | $2,450 |
| Claude Fable 5.1Frontier | $0.245 | $4,900 |
| GPT-6 AstraFrontier | $0.245 | $4,900 |
How it calculates
An agent resends its whole on every turn. That is why tokens in, not the answer out, drive the bill. The calculator assumes the fixed part, system prompt and tool definitions, is sent on every turn, and that the history grows by the same amount each turn.
The defaults are the worked example from the article on what an AI agent costs: 20,000 requests, 4 turns and 3,000 tokens of fixed context. That comes to 18,500 tokens in and 1,200 out per request.
The prices
Dollars per million , retrieved 14 September 2026.
| Model | Tier | Input | Output |
|---|---|---|---|
| Claude Fable 5.1 | Frontier | $10 | $50 |
| GPT-6 Astra | Frontier | $10 | $50 |
| Claude Opus 5 | Frontier | $5 | $25 |
| GPT-5.6 Terra | Mid tier | $2 | $12 |
| Gemini 3.1 Pro (preview) | Mid tier | $2 | $12 |
| Claude Sonnet 5 | Mid tier | $2 | $10 |
| Mistral Medium 3.5 | Mid tier | $1.5 | $7.5 |
| Grok 4.6 | Mid tier | $2 | $6 |
| Claude Haiku 4.5 | Small | $1 | $5 |
| Gemini 3.8 Flash | Small | $0.75 | $3.75 |
| Gemini 3.5 Flash Lite | Small | $0.30 | $2.5 |
| GPT-5.6 Luna | Small | $0.20 | $1.2 |
| Mistral Small | Small | $0.15 | $0.60 |
A model is only included when OpenRouter and LiteLLM report exactly the same price. Prices change often, and a single source has been wrong before.
What we could not confirm
These models are left out because the sources disagree, or because one of them has no price. We show both numbers rather than picking one.
| Model | OpenRouter | LiteLLM |
|---|---|---|
| GPT-5.6 Sol | $2 / $10 | $4 / $20 |
| Kimi K3 | $2.648 / $13.283 | $3 / $15 |
| DeepSeek V4 Pro | $1.6 / $3.2 | $1.32 / $3.96 |
| DeepSeek V4 Flash | $0.079 / $0.159 | $0.30 / $1.2 |
What the calculator leaves out
- Caching. Providers discount the fixed part of the context when it is . That can cut the bill sharply, but the discount varies by provider and is not cross-checked here.
- Batch. Several providers charge less for jobs that do not need an answer straight away.
- Failures and retries. Calls that fail and run again cost money too.
- Currency and taxes. Prices are list prices in dollars. Exchange rates and any VAT come on top.
- Different tokenizers. The same text does not become the same number of tokens everywhere. Anthropic says Claude 4.7 and later produce about 30 percent more tokens for the same text, so the same numbers in the fields above do not mean exactly the same amount of text for every model.
- Introductory prices. Gemini 3.8 Flash is on an introductory price through 2026. From 1 January 2027 the price doubles, according to Google.