AI model cost calculator:
14 models compared in euro.
Enter the input and output tokens for one task and the number of tasks a month. The calculator shows the monthly cost of 14 current models from OpenAI, Anthropic, Google, Mistral, Together AI and DeepSeek, lowest first, in euro. Prices are list prices from each vendor’s pricing page, checked on 19 November 2026.
How AI models charge
Model APIs charge per token. A token is a piece of a word: about 750 English words make 1,000 tokens. Vendors set one price for input tokens (what you send: instructions, documents, conversation history) and a higher price for output tokens (what the model writes). Prices are per million tokens.
The formula for one month is: tasks × (input tokens × input price + output tokens × output price) ÷ 1,000,000. The calculator then converts US dollars to euro at the ECB reference rate of 19 November 2026: US$1.0992 to €1.
Reasoning models count their internal thinking as output tokens. If a model thinks before it answers, the output count per task can be several times the length of the visible answer. Measure it on a real test before you trust an estimate.
Compare model costs for your task
Untick models to remove them from the list. The results sort from lowest to highest cost.
What one month of tasks costs.
List prices per million tokens from each vendor’s pricing page, checked 19 November 2026. We convert US dollars to euro at the ECB reference rate of 19 November 2026 (US$1.0992 to €1). Prices exclude VAT.
Estimate at list prices. Most vendors cut the input price for cached prompts and offer about half price for batch jobs that can wait, so real bills are often lower. Reasoning models also bill their thinking as output tokens. The model comparison covers quality, data location and terms.
The prices the calculator uses
| Model | Input | Output | Cached input |
|---|---|---|---|
| OpenAI GPT-6.2 Astra | 8.00 | 40.00 | 0.40 |
| OpenAI GPT-6.2 Sol | 2.00 | 10.00 | 0.10 |
| OpenAI GPT-6.1 Terra | 1.00 | 5.00 | 0.05 |
| OpenAI GPT-6 Luna | 0.10 | 0.50 | 0.01 |
| Anthropic Claude Opus 5.5 | 4.00 | 20.00 | 0.20 |
| Anthropic Claude Sonnet 5.5 | 2.00 | 10.00 | 0.10 |
| Anthropic Claude Haiku 5.5 (prompts up to 100k) | 0.10 | 0.50 | 0.01 |
| Google Gemini 3.1 Pro, preview (prompts up to 200k) | 2.00 | 12.00 | 0.20 |
| Google Gemini 3.8 Flash | 0.75 | 3.75 | 0.075 |
| Google Gemini 3.5 Flash-Lite | 0.30 | 2.50 | 0.03 |
| Mistral Large 4 | 0.70 | 2.10 | 0.07 |
| Mistral Small 4 | 0.15 | 0.60 | 0.015 |
| gpt-oss-120B on Together AI (open weights) | 0.15 | 0.60 | not listed |
| DeepSeek V4.1 Flash (peak hours) | 0.30 | 1.20 | 0.006 |
Notes on the table: Gemini 3.1 Pro is a preview model. Google has announced that Gemini 3.8 Flash rises to US$1.50 input and US$7.50 output on 1 January 2027. Mistral’s FAQ quotes a lower price for "Mistral Large" without a version; we use the versioned price from Mistral’s documentation. DeepSeek charges half price outside its peak hours. gpt-oss-120B is published by OpenAI under the Apache 2.0 licence, so you can also run it on your own servers.
A worked example: 10,000 tasks a month
The calculator starts at 2,000 input tokens and 500 output tokens per task, 10,000 tasks a month. That is roughly the size of a short email triage or a one-page document summary. It is an example, not a benchmark.
| Model | Per month |
|---|---|
| GPT-6 Luna or Claude Haiku 5.5 | €4.09 |
| Mistral Small 4 or gpt-oss-120B | €5.46 |
| DeepSeek V4.1 Flash (peak) | €10.92 |
| Gemini 3.5 Flash-Lite | €16.83 |
| Mistral Large 4 | €22.29 |
| Gemini 3.8 Flash | €30.70 |
| GPT-6.1 Terra | €40.94 |
| GPT-6.2 Sol or Claude Sonnet 5.5 | €81.88 |
| Gemini 3.1 Pro (preview) | €90.98 |
| Claude Opus 5.5 | €163.76 |
| GPT-6.2 Astra | €327.51 |
The most expensive model costs 80 times the cheapest for the same task. That does not mean the cheapest is the right choice. Run 50 to 100 of your real tasks through two or three candidates and check the answers. Pick the cheapest model that passes your test. Our model comparison covers quality, data location and terms, which matter as much as price.
How many tokens does a task use?
The token count depends on what you send. Use these rules of thumb to start, then measure your own tasks:
- Fixed instructions: every request includes the system instructions. A detailed instruction set of 1,500 words is about 2,000 tokens on every call, which is why caching helps.
- Documents: count the words in the document and multiply by about 1.33 (the common rule of thumb of 750 words to 1,000 tokens). A 10-page contract of about 5,000 words is roughly 6,700 tokens.
- Conversations: each new turn sends the whole conversation so far. A 10-turn chat costs much more than 10 single questions.
- Tool results: when an agent looks up a booking or a customer record, the result goes back into the input.
- Output: a short reply is 50 to 200 tokens. A one-page summary is about 500 to 700. Thinking tokens on reasoning models come on top.
Words-to-tokens ratios differ by language and by model, and the ratios above are for English text. If your customers write in several languages, test each one.
A practical method: run 20 real tasks through the API, read the input and output token counts that the API returns, and take the average. Put those averages into the calculator. Repeat after launch, because real use is often longer than tests.
Caching and batch discounts
The calculator uses standard list prices. Two discounts can cut real bills a lot:
- Prompt caching. When many requests start with the same text (instructions, a policy document), the vendor charges less for the repeated part. Anthropic charges 0.05 times the base input price for a cache hit on Opus 5.5 and Sonnet 5.5, and 0.1 times on other models; writing to the cache costs more. Mistral says cached input cuts input cost by up to 90%. Google also charges for cache storage per hour.
- Batch processing. If results can wait (overnight reports, bulk document checks), OpenAI, Anthropic, Google and Mistral offer about 50% off. DeepSeek has no batch discount but charges half price off-peak.
An agent that answers customers in real time cannot use batch, but it can use caching. A nightly job that summarises the day’s emails can use both. See the AI agent cost calculator for the full monthly cost of an agent, including WhatsApp and phone.
What to check beyond the price
- Data location. Some vendors offer EU processing; some do not. Ask each vendor where it processes and stores your requests. Read EU data residency for AI models before you send personal data.
- Terms. Check that the vendor does not train on your API data, and that a data processing agreement is in place.
- Stability. Preview models can change or be withdrawn. Mistral deprecated Medium 3.5 on 31 October 2026.
- Speed and rate limits. A cheap model with low rate limits can slow a busy agent.
The OpenAI and Anthropic integration pages describe what each API can do in an agent.
Common questions
How do I find the token count for my task?
Run a few real examples through the vendor’s API or playground and read the usage numbers in the response. Most APIs return input and output token counts with every request. Average them over at least 20 examples. Include the instructions and any documents, not only the user’s question.
Why is the price in US dollars converted to euro?
All the vendors in the table list API prices in US dollars. We convert at the ECB reference rate of 19 November 2026 (US$1.0992 to €1) so you can compare with other euro costs. Your real cost depends on the rate when you pay and any card charges.
Is an open-weight model always cheaper?
Not always. gpt-oss-120B on Together AI is among the cheapest in the table, but the smallest models from OpenAI and Anthropic cost less at list price. Running an open model on your own servers removes per-token fees but adds hardware, hosting and staff time. It makes sense when data control matters more than cost.
How often do these prices change?
Often. Vendors release new models several times a year and sometimes change prices of existing ones. Google has already announced a price rise for Gemini 3.8 Flash on 1 January 2027. Check the vendor’s pricing page before you commit, and review your model choice every few months.
Sources
- OpenAI: API pricing
- Anthropic: Claude API pricing
- Google: Gemini API pricing
- Mistral: API pricing (docs)
- Mistral: pricing and FAQ
- Together AI: pricing
- Hugging Face: openai/gpt-oss-120b model card (licence)
- DeepSeek: API pricing
- European Central Bank: euro reference rates
We last checked these sources on 19 November 2026. Rules, fees and prices change. Check the source before you act.
Bring us the task.
Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.
Start a project.

Book a call or send us a message. We read everything.
tash@neu.ie
Working with teams in Ireland and worldwide.
Free 30-minute AI review.